Index

LIGHT FLICKER EVALUATION OF ELECTRIC ARC FURNACES BASED
ON NOVEL SIGNAL PROCESSING ALGORITHMS

A THESIS SUBMITTED TO
THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES
OF
MIDDLE EAST TECHNICAL UNIVERSITY
BY
NESL
IHAN K
OSE
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR
THE DEGREE OF MASTER OF SCIENCE
IN
ELECTRICAL AND ELECTRONICS ENGINEERING
SEPTEMBER 2009
Approval of the thesis:
LIGHT FLICKER EVALUATION OF ELECTRIC ARC FURNACES
BASED ON NOVEL SIGNAL PROCESSING ALGORITHMS
submitted by NESL
IHAN K
OSE in partial fulllment of the requirements for

the degree of Master of Science in Electrical and Electronics Engineer-
ing Department, Middle East Technical University by,
Prof. Dr. Canan

Ozgen
Dean, Graduate School of Natural and Applied Sciences
Prof. Dr.

Ismet Erkmen
Head of Department, Electrical and Electronics Engi-
neering
Prof. Dr. Kemal Leblebicioglu
Supervisor, Electrical and Electronics Engineering De-
partment, METU
Dr.

Ozg ul Salor
Co-supervisor, T
UB
ITAK-UZAY; Electrical and Elec-

tronics Engineering Department, Gazi

Univ.
Examining Committee Members:
Prof. Dr. Muammer Ermis
Electrical and Electronics Engineering, METU
Prof. Dr. Kemal Leblebicioglu
Assoc. Prof. Dr. Tolga C iloglu
Prof. Dr. Isk C adrc
Electrical and Electronics Engineering, Hacettepe University
Assist. Prof. Dr. C agatay Candan
Date: 10 / 09 / 2009
I hereby declare that all information in this document has been ob-
tained and presented in accordance with academic rules and ethical
conduct. I also declare that, as required by these rules and conduct,
I have fully cited and referenced all material and results that are not
original to this work.
Name, Last Name: NESL
IHAN K
OSE
Signature :
iii
ABSTRACT
LIGHT FLICKER EVALUATION OF ELECTRIC ARC FURNACES BASED
ON NOVEL SIGNAL PROCESSING ALGORITHMS
Kose, Neslihan
M.S., Department of Electrical and Electronics Engineering
Supervisor : Prof. Dr. Kemal Leblebicioglu
Co-Supervisor : Dr.

Ozg ul Salor
September 2009, 81 pages
In this research work, two new ickermeters are proposed to estimate the light
icker caused by electric arc furnaces (EAFs) where the system frequency devi-
ates signicantly. In these methods, analytical expressions of the instantaneous
light icker sensation are obtained beginning from a voltage waveform and these
expressions are used to obtain a icker estimation method based on the IEC (In-
ternational Electrotechnical Commission) ickermeter. First method is a spec-
tral decomposition based approach using DFT to estimate the light icker. The
leakage eect of the DFT algorithm due to fundamental frequency variation
is reduced by employing spectral amplitude correction procedure around the
fundamental frequency. Second method is a Kalman lter based approach, in
which the frequency domain components of the voltage waveform are obtained
by Kalman ltering. Then these components are used to obtain the light icker.
Since the frequency decomposition is obtained by Kalman ltering, no leakage
eect of the DFT is involved in case of frequency deviations which is an impor-
tant advantage. Both methods are tested on both simulated data and eld data
iv
obtained from three dierent EAF plants where the icker level and frequency
variation is considerably high. The comparison with the digital realization of
the IEC ickermeter shows that the methods are successful in estimating light
icker with low computational complexity. The methods are especially useful
for conditions such as disturbances and subsequent system transients where the
system frequency deviates signicantly, since the methods avoid the need for
online sampling rate adjustment to prevent the DFT leakage eect.
Keywords: icker, electrical arc furnace (EAF), power quality (PQ), Kalman
lter (KF), DFT leakage
v
OZ
OZG
UN S
INYAL

ISLEME ALGOR
ITMALARI KULLANARAK ELEKTR
IK
ARK OCAKLARINDA ISIK KIRPISMASININ DE
GERLEND
IR
ILMES
I
Kose, Neslihan
Y uksek Lisans, Elektrik ve Elektronik M uhendisli gi Bol um u
Tez Yoneticisi : Prof. Dr. Kemal Leblebicioglu
Ortak Tez Yoneticisi : Dr.

Ozg ul Salor
Eyl ul 2009, 81 sayfa
Bu arastrma calsmasnda, sistem frekansnn saptg durumlarda elektrik ark
ocaklarnda (EAO) krpsma degerini hesaplamaya yonelik iki yeni krpsma
olcer onerilmistir. Bu yontemlerde, gerilim dalgasn anlk sk krpsma algsna
baglayan analitik ifadelere dayanarak, IEC krpsma olcer standardna uygun
bir krpsma kestirim yontemi uygulanmstr.

Ilk yontem, spektral ayrstrma
yontemine dayal ve Ayrk Fourier Don us um un un (DFT) kullanldg bir krpsma
hesaplama yaklasmdr. Bu yontemde, temel frekanstaki degisim sebebiyle
DFT algoritmasnda olusan kacak etki, temel frekans etrafnda spektral genlik
d uzeltme yontemi uygulayarak azaltlmaktadr.

Ikincisi ise, Kalman s uzgeci kul-
lanmna dayal bir sk krpsmas hesaplama yontemidir. Bu yontemde, Kalman
s uzgeci kullanarak dalgann frekans spektrumundaki bilesenleri elde edilmekte
ve bu bilesenler sk krpsmasnn hesaplanmas icin kullanlmaktadr. Frekans
ayrsm Kalman s uzgecleme ile elde edildiginden, frekans sapmas oldugunda
DFTde olusan kacak etki gozlenmemektedir; bu da yontemin sundugu onemli
bir avantajdr. Her iki yontem de, hem sentetik veriler, hem de krpsma dere-
vi
cesinin ve frekans degisiminin oldukca y uksek oldugu uc farkl EAO fabrikasndan
toplanan saha verileri uzerinde test edilmistir. IEC krpsma olcerin saysal
uyarlamasyla yaplan karslastrmalar, yontemlerin d us uk hesaplama y uk u ile
sk krpsmas tespitinde basarl sonu clar verdigini gostermektedir. Yontemler,
ozellikle sistem frekansnn kaydg bozucu etkenlerin varlgnda ve bunu izleyen
sistem gecici rejimleri gibi durumlar icin uygundur, c unk u DFTnin ka cak etk-
isini onlemek icin uygulanan cevrimici ornekleme frekans ayar gereksinimi or-
tadan kaldrmaktadrlar.
Anahtar Kelimeler: krpsma, elektrik ark ocag (EAO), g uc kalitesi (GK),
Kalman S uzgeci (KS), DFT kacak etkisi
vii
To My Family
viii
ACKNOWLEDGMENTS
I would like to express my deepest gratitude to my supervisor, Prof. Dr. Ke-
mal Leblebicioglu, for his guidance, advices, boundless knowledge transfer and
criticism throughout this research work. I would like to thank my supervisor
especially for his valuable contributions to my career.
I also would like to thank my co-supervisor, Dr.

Ozg ul Salor, not only for her
contributions and suggestions throughout this research, but also for her contin-
uous condence in me. Her encouragement and continuous support throughout
this research work are the main reasons for the success of the study. I express my
sincerest thanks to her for her guidance, companionship, support and valuable
contributions during my graduate studies.
I wish to express my special thanks to Prof. Dr. Muammer Ermis for his
guidance, encouragement and valuable contributions to my career.
I would like to thank Prof. Dr. Isk C adrc for her support, comments and
suggestions during my graduate studies.
This research work is carried out through the National Power Quality Project
of Turkey in T
UB
ITAK-UZAY Power Electronics Group. I would like to thank

the Public Research Institutions and Development Projects Support Group (KA-
MAG) of the T
UB
ITAK for full nancial support of the project.

Special thanks to mobile power quality measurement team of National Power
Quality Project of Turkey for obtaining the eld data which I have used in the
research work.
I would like to acknowledge my friends in Power Electronics Group of T
UB
ITAK
UZAY for their companionship and also for the sincere job environment provided
by them.
ix
I also would like to thank all other my friends who have been with me during
this period. Their friendship, their encouragement and their trust in me are
always very important for me.
Finally, I would like to thank my dearest family, my father S ukr u, my mother
Birsen and my sister Burcu for their continuous support, but especially for their
trust in me that I could accomplish this task during this period. I always feel
lucky to have such a lovely family, I love you all...
x
TABLE OF CONTENTS
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
OZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . ix
TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . . xi
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi
CHAPTERS
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Denition and Scope of the Thesis . . . . . . . 2
2 A NEW SPECTRAL DECOMPOSITION BASED APPROACH
FOR FLICKER EVALUATION OF ELECTRIC ARC FURNACES 7
2.1 The IEC Flickermeter . . . . . . . . . . . . . . . . . . . 7
2.2 Derivation of the Mathematical Relationship between the
Input Voltage and the Flicker Sensation . . . . . . . . . 9
2.2.1 Square Demodulation and Lamp-Eye-Brain Fil-
ters . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Squaring Multiplier and Sliding Mean Filter . . 11
2.2.3 Instantaneous Flicker Sensation Determination
and the Proposed Method . . . . . . . . . . . . 12
2.3 Correction of the Spectral Components for Deviated Fre-
quency Case . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Formulation of Leakage Eect of the Discrete
Fourier Transform . . . . . . . . . . . . . . . . 14
xi
2.3.2 Correction of the Spectral Components . . . . 16
2.4 Verication of the Proposed Method . . . . . . . . . . . 26
2.4.1 Verication with Simulated Data . . . . . . . . 26
2.4.1.1 Test 1: Single Flicker Frequency Case 26
2.4.1.2 Test 2: Rectangular Voltage Fluctuations-
Combinations of Flicker Frequencies 27
2.4.1.3 Test 3: Homogeneity of the Flicker-
meter Response . . . . . . . . . . . 27
2.4.2 Verication on AC Arc Furnace Data . . . . . 28
3 A KALMAN FILTERING BASED APPROACH FOR LIGHT
FLICKER EVALUATION OF POWER SYSTEMS . . . . . . . 36
3.1 General Characteristics of Kalman Filters . . . . . . . . 36
3.2 Obtaining the Spectral Decomposition based on Kalman
Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.1 Verication of Kalman Filtering Method for In-
terharmonic Detection . . . . . . . . . . . . . . 47
3.2.1.1 Analysis on Simulative Data . . . . 47
3.2.1.2 Analysis on Field Data . . . . . . . 50
3.2.2 Discussion . . . . . . . . . . . . . . . . . . . . 52
3.3 Flicker Evaluation Based on Kalman Filtering . . . . . . 54
3.4 Verication of the Proposed Method for Flicker Analysis 59
3.4.1 Verication with Simulated Data . . . . . . . . 59
3.4.1.1 Test 1: Single Flicker Frequency Case 60
3.4.1.2 Test 2: Combinations of Dierent
Flicker Frequencies Case . . . . . . 60
3.4.1.3 Test 3: Homogeneity of the Flicker-
meter Response . . . . . . . . . . . 61
3.4.2 Verication on AC Arc Furnace Data . . . . . 62
4 CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . 71
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
APPENDICES
xii
A IEC FLICKERMETER RESPONSE FOR RECTANGULAR FLUC-
TUATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
xiii
LIST OF TABLES
TABLES
Table 2.1 Normalized Flickermeter Response for Sinusoidal Voltage Fluc-
tuations, 230V/50Hz system - IEC Standard 61000-4-15 . . . . . . . 13
Table 2.2 Eect of the System Frequency Deviation . . . . . . . . . . . 17
Table 2.3 Proposed light icker estimation algorithm response for sinu-
soidal voltage uctuations . . . . . . . . . . . . . . . . . . . . . . . . 27
Table 2.4 Proposed icker estimation algorithm response for rectangular
voltage uctuations . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Table 2.5 Homogeneity of the proposed light icker estimation algorithm 29
Table 2.6 Short term icker severity, P
st
, computed inside the intervals
given in Fig 2.12 with the proposed method and the digital realization
of the IEC ickermeter . . . . . . . . . . . . . . . . . . . . . . . . . 30
st
given in Figure 2.13 with the proposed method and the digital real-
ization of the IEC ickermeter . . . . . . . . . . . . . . . . . . . . . 31
st
given in Figure 2.14 with the proposed method and the digital real-
ization of the IEC ickermeter . . . . . . . . . . . . . . . . . . . . . 32
Table 2.9 Comparison of the short term icker severity, P
st
, computed
inside the intervals given in Fig 2.12 with the proposed method, the
method given in [5] and the digital realization of the IEC ickermeter 35
Table 3.1 Both group and subgroup interharmonic-1 computed with the
standard FFT computation (FFT), the method in [23] and the pro-
posed method (PM) for x(t) = sin(2ft) + 0.5sin(265t) . . . . . . 49
xiv
Table 3.2 Both group and subgroup interharmonic-1 computed with the
standard FFT computation (FFT), the method in [23] and the pro-
posed method (PM) for x(t) = sin(2ft)+0.7sin(253t)+0.6sin(255t)+
0.5sin(258t) + 0.4sin(262t) + 0.3sin(265t) . . . . . . . . . . . . 50
Table 3.3 Group computation of interharmonic-1 with the proposed method,
the method in [23], the standard FFT computation (FFT) and the
FFT+interpolation method on eld data. Error values are computed
assuming that the FFT+interpolation gives reference results. . . . . 52
Table 3.4 Proposed light icker estimation algorithm response for sinu-
soidal voltage uctuations . . . . . . . . . . . . . . . . . . . . . . . . 60
Table 3.5 Homogeneity of the proposed light icker estimation algorithm 62
st
given in Fig 3.8 with the proposed method, method in [7] and the
digital realization of the IEC ickermeter . . . . . . . . . . . . . . . 64
st
st
Table 4.1 The Eect of Changing the Number of Samples per Window to
Instantaneous Flicker Sensation, S . . . . . . . . . . . . . . . . . . . 76
Table A.1 Normalized Flickermeter Response for Rectangular Voltage Fluc-
tuations, 230V/50Hz system - IEC Standard 61000-4-15 . . . . . . . 81
xv
LIST OF FIGURES
FIGURES
Figure 1.1 An example showing harmonics and interharmonics of an AC
arc furnace data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Figure 2.1 Block diagram of the IEC ickermeter. . . . . . . . . . . . . 8
Figure 2.2 An example waveform showing the icker-causing uctuation
of the voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Figure 2.3 The leakage eect for frequency deviations on the instanta-
neous icker sensation, S. Fundamental frequency is changed from
49.9Hz to 50.1Hz with 0.01Hz increments. 200ms-length window is
used for DFT computation. . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 2.4 DFT leakage eect on the amplitude characteristics for the
fundamental frequency values given in Table 2.2. . . . . . . . . . . . 19
Figure 2.5 S computed out of the DFT values given in Figure 2.4 for a
data of 10sec. long without spectral correction. . . . . . . . . . . . . 19
Figure 2.6 Block diagram of the proposed icker estimation method. . . 20
Figure 2.7 Zero-crossing correction [23]. . . . . . . . . . . . . . . . . . . 21
Figure 2.8 Amplitude of X
w
(). Xs and Os show the frequency samples
without and with frequency deviation, respectively [23]. . . . . . . . 22
Figure 2.9 Spectral amplitude correction. . . . . . . . . . . . . . . . . . 24
Figure 2.10 Amplitude correction of the 55Hz component in Figure 2.9.
Correction is achieved by vector subtraction of the synthetic compo-
nent from the original component using both phases and amplitudes
given in Figure 2.9. . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
xvi
Figure 2.11 Eect of frequency variation on estimation of S. (a) Frequency
variation, (b) estimation of S without spectral correction compared
with the IEC ickermeter result, (c) estimation of S with the proposed
method (with spectral correction) compared with the IEC ickermeter
result. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Figure 2.12 Plant-1 (a) rms voltage at the MV side of the EAF Plant with
the P
st
computed intervals, (b) fundamental frequency variation, (c)
instantaneous icker sensation, S. . . . . . . . . . . . . . . . . . . . 30
the P
st
the P
st
Figure 2.15 Plant-1 instantaneous icker sensation, S, for Interval-1 and
Interval-2 in Figure 2.12 for a more detailed comparison of the pro-
posed method and the IEC ickermeter. . . . . . . . . . . . . . . . . 33
Figure 3.1 Block diagram of the operation of the Kalman Filter [27]. . . 39
Figure 3.2 Block diagram of the operation of the Extended Kalman Filter
[27]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Figure 3.3 Block diagram of the proposed spectral decomposition method. 41
Figure 3.4 Voltage waveform containing dierent interharmonic frequencies. 42
xvii
Figure 3.5 Magnitude and the fundamental frequency variation of the
eld data collected at the MV side of the power system supplying an
arc furnace plant. Fundamental frequency deviation is obtained by
using EKF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Figure 3.6 Block diagram of the proposed icker estimation method. . . 55
Figure 3.7 Simulated voltage waveform, v(t), and the estimated envelope
of v(t) which is obtained by the EKF. . . . . . . . . . . . . . . . . . 61
the P
st
the P
st
the P
st
Interval-2 in Figure 3.8 for a more detailed comparison of the proposed
method and the IEC ickermeter. . . . . . . . . . . . . . . . . . . . 67
Interval-2 in Figure 3.9 for a more detailed comparison of the proposed
method and the IEC ickermeter. . . . . . . . . . . . . . . . . . . . 67
Figure 3.14 Histogram of fundamental frequency of Plant-1. . . . . . . . 69
xviii
CHAPTER 1
INTRODUCTION
1.1 General
Light icker resulting from frequently recurring, low amplitude voltage uctua-
tions is one of the major power quality disturbances in a power system. Light
icker is due to the uctuation of the voltage amplitude supplying the incandes-
cent lamps inside such a frequency and amplitude range that it causes annoying
eect on human-beings. Such a uctuation is usually caused by variable loads
in the medium or high (MV or HV) voltage side such as arc furnaces and ladle
furnaces, or by variable loads on the LV side such as X-ray equipment or copy-
ing machines. Therefore, it is a fact that voltage uctuations caused by rapidly
changing loads in the power system may cause noticeable light icker. Up to
now, researches have shown that human eye is sensitive to voltage uctuations
below 50Hz. Especially voltage uctuations between 0.5Hz and 25Hz cause
annoying eect on human beings.
IEC (International Electrotechnical Commission) has standardized the light
icker measurement, which employs weighting curves that simulate the human
eye-brain response to light uctuations caused by an incandescent lamp [1]. The
IEC standard assumes a continuous voltage input and outputs a continuous in-
stantaneous icker sensation, and a discrete output, which is called the short
term icker severity. On the other hand, several approaches have been used to
estimate the instantaneous icker sensation and the icker severity when the
input is a sampled form of the continuous voltage [2]-[9]. Most of these ap-
1
proaches use the frequency domain decomposition of the voltage waveform to
obtain the instantaneous icker sensation [2]-[6]. The problem with frequency
domain approaches is that they usually employ DFT to obtain the frequency
decomposition, which suers from picket-fence eect and leakage due to the fun-
damental frequency deviations and the windowing eect in the digital domain
[4]. To solve this problem, various approaches have been proposed including
indirect demodulation method [6], smart DFT [2] and wavelet transforms [3].
In indirect demodulation method, initially the envelope of the voltage waveform
is obtained and then DFT algorithm is applied on this envelope. Therefore,
no leakage eect problem due to fundamental frequency deviation occurs [6].
Smart DFT is also another approach to solve the leakage problems. Smart DFT
smartly avoids the errors that arise when fundamental frequency deviates and
keeps the advantages of the DFT algorithm [2]. In [3], a wavelet transform ap-
proach is proposed to estimate the light icker, which is considerably robust to
fundamental frequency deviations. These approaches are usually compared to
the digital realizations of the IEC ickermeter [9] and quite satisfactory results
are obtained at the expense of computational complexity.
1.2 Problem Denition and Scope of the Thesis
Up to now, researches have shown that the system frequency deviates signif-
icantly due to the EAF operation. Power quality measurements realized na-
tionwide have shown that the 50Hz system frequency deviates within a band
of 500.2Hz at the MV side of the busbar supplying the EAFs [10]. It has
been reported in the literature that the actual system frequency may deviate
from the normal value due to various reasons such as disturbances, subsequent
transients and or the dramatic real power uctuations of arc furnace plants
[3, 11, 12, 13, 14]. In [3], it is reported that eld tests performed in a steel
plant in southern Taiwan, it was observed that the frequency of R-phase voltage
measurement changes from 59.92Hz to 60.12Hz, where the nominal frequency
is 60.0Hz. In [11], it is reported that the most preferable frequency deviation is
0.02Hz; however, for case studies such as a large steel mill and an arc furnace
2
factory, the maximum allowable change in the transient frequency lies between
0.50Hz. In [15], it is reported that US-wide power system frequency usually
varies between 59.99Hz and 60.02Hz;, however, in case of an event the frequency
may drop down to 59.55Hz. It is observed that during safe operation, the power
system frequency deviates due to various reasons; but, in interconnected systems
such as the US-power network [15], this deviation is much less than that of the
power systems without such an interconnection [10, 11]. If a DFT-based algo-
rithm is employed to obtain spectral decomposition of a waveform in a deviated
frequency case, leakage eect occurs and to overcome the leakage eect, generally
two methods can be used. The rst one is to adjust the sampling frequency such
that the same number of samples cover the same number of frequency cycles.
This can be achieved, for example, using a phase-locked-loop controlling the
sampling process. The second one is to use a xed-length window with dierent
spectral techniques to obtain a better frequency spectrum estimate [2, 3, 6]. In
this research work, two new methods are proposed to estimate spectral compo-
nents of voltage waveforms and consequently the light icker where the system
frequency deviates signicantly due to the EAF operation. In the proposed
methods, after obtaining the spectral components, the light icker caused by
electric arc furnace (EAF) plants at the MV side is calculated by using these
spectral components.
Harmonics and interharmonics are frequently observed in power systems. Har-
monic frequency is the frequency which is an integer multiple of the fundamental
frequency (50Hz or 60Hz), whereas interharmonic frequency is the frequency
which is not an integer multiple of the fundamental frequency [16]. Several
interharmonics and harmonics are observed on the data collected from EAF
plants. Spectral analysis of harmonics and interharmonics of the power signals
are usually done by using the Fast Fourier Transform (FFT) algorithm due to its
computational complexity. Figure 1.1 shows the harmonics (up to 5. harmonic,
which is 250Hz for 50Hz power system) and interharmonics of a 200ms-length
eld data which are obtained by applying FFT algorithm on the voltage data
of an EAF plant. Interharmonics are the real cause of light icker in power sys-
tems. Flicker-causing interharmonics can be observed clearly in the frequency
3
spectrum of the specied voltage waveform as shown in Figure 1.1.
0 50 100 150 200 250
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Freq (Hz)
V
FFT Amplitudes of the Original Waveform
50 Hertz amplitude (normalized to 100V)
Figure 1.1: An example showing harmonics and interharmonics of an AC arc
furnace data.
In this research work, both of the proposed methods are spectral decomposition
based methods to obtain icker-causing interharmonics. The models proposed
in both of these methods are based on the IEC ickermeter explained in the
IEC Standard 61000-4-15 [1], whose block diagram is given in Figure 2.1. Both
proposed methods simulate the IEC ickermeter computation beginning from
the normalized voltage input (input to Block - 2 in Figure 2.1) to Output-
5 which outputs S in digital domain. In these methods, a frequency domain
approach is proposed to estimate the light icker. The main dierence between
the two proposed methods is that; to obtain the frequency spectrum of the
voltage waveform, considerably dierent approaches are applied. In the rst
method, spectral decomposition is obtained by utilizing the DFT algorithm on
the voltage waveform using 200ms time windows (10 cycles for a 50Hz system),
which coincides with the IEC harmonic and interharmonic computation standard
[16]. However due to the system frequency deviation, leakage eect is observed
when DFT algorithm is applied. The leakage eect of the DFT algorithm due
4
to frequency variation is taken care of by employing a novel spectral amplitude
correction procedure around the fundamental frequency [7, 17]. The correction
procedure is based on estimating the system frequency at consecutive windows
and synthesizing a pure sinusoidal signal with the same frequency so as to obtain
the leakage eect caused by the system frequency deviation. The leakage is
then vector-subtracted for each interharmonic around the fundamental frequency
and corrected frequency spectrum is obtained. In the second method, on the
other hand, spectral decomposition is obtained by applying two Kalman lters
consecutively in succession on the voltage waveform again using 200ms time
windows as recommended in the IEC harmonic and interharmonic computation
standard [16]. Initially, an EKF is used to obtain the voltage envelope and
then a second Kalman lter, which is a LKF this time, is used to obtain the
spectral components around the fundamental out of the envelope of the voltage
waveform causing icker [18]. After the spectral components are obtained by the
Kalman lters, the light icker is evaluated using these components as proposed
in the rst method, which is also explained in [7, 19]. Since in the second
method, no DFT algorithm is used, no leakage eect problem is involved in case
of fundamental frequency deviations. Therefore there is no need for extra work
to obtain a corrected spectrum, which is an important advantage and novelty of
the second method.
In both of the proposed methods, the eye-brain weighting curve of the IEC
ickermeter is realized by comparing the voltage spectrum with the tabulated
normalized ickermeter responses for sinusoidal voltage uctuations of the IEC
standard [1]. Both of the proposed methods are tested on both simulated data
and eld data obtained from three dierent EAF plants. The comparison with
the digital realization of the IEC ickermeter shows that the proposed methods
give satisfactory estimations of both the instantaneous icker sensation and the
short term icker severity with low computational complexity. Therefore, the
proposed methods can be considered as two new ickermeter applications.
This thesis is organized as follows: Chapter 2 gives a brief overview of the IEC
icker determination approach applied to a discrete voltage waveform, on which
the proposed methods are based. The rst proposed method, a new spectral
5
decomposition based approach for icker evaluation of electric arc furnaces, with
algorithm improvements are also described in this chapter. In Chapter 3, the
second proposed method, a Kalman Filtering Based Approach for Light Flicker
Evaluation of Power Systems is presented. The comparison between the two
proposed methods on AC arc furnace data is also given in this chapter. Finally
in Chapter 4, conclusions and future directions are provided.
6
CHAPTER 2
A NEW SPECTRAL DECOMPOSITION BASED
APPROACH FOR FLICKER EVALUATION OF
ELECTRIC ARC FURNACES
The proposed spectral decomposition based approach, which is also explained
in [7], is attributed to the mathematical relationship derived between the input
waveform of the IEC ickermeter and the measured instantaneous icker sensa-
tion. Therefore, this chapter briey explains the IEC ickermeter design rst,
then it presents the mathematical derivations followed by the proposed method.
2.1 The IEC Flickermeter
Both of the methods proposed in this thesis are based on the IEC ickerme-
ter explained in the IEC Standard 61000-4-15 [1]. The IEC ickermeter block
diagram is given in Figure 2.1. The major parts of the IEC ickermeter are
the input processing part, the lamp-eye-brain response part, and the output
processing part. There are two primary steps in input processing part. First
step is for converting the rms value of measured voltage to a reference level in
order to guarantee that percentage deviations are equal regardless of the input
rms level. In second step, the square of the input is obtained by the square
demodulator to separate the low frequency (0.525Hz) components from the
power frequency components through ltering. Then comes the lamp-eye-brain
response part. In this part, the lamp-eye-brain characteristic is obtained from
mathematical derivation of the response of a lamp to a supply voltage variation,
7
Figure 2.1: Block diagram of the IEC ickermeter.
the perception ability of the human eye and the memory tendency of a human
brain. Finally in output processing part, the output of Block 4, instantaneous
icker sensation is translated into the statistical index, P
st
. Therefore, the IEC
ickermeter can be basically explained as; the IEC procedure for light icker
computation initially scales the input voltage down to the last minute rms value
before it simulates the response of the lamp-eye-brain chain by a square law
demodulator followed by weighting lters and a range selector. Then follows
the squaring multiplier and the sliding mean lter blocks, which simulate the
non-linear eye-brain perception and the storage eect of the human brain, re-
spectively. The result gives the instantaneous icker sensation, denoted by S,
which is a continuous waveform (Output - 5 in Figure 2.1). Then S is sampled
and put into a statistical evaluation block to obtain the short term icker sever-
ity, P
st
, which is computed for non-overlapping data windows of 1, 2, 10 or 15
minutes. The short term icker severity index, P
st
, is a statistical quantication
of the instantaneous icker sensation, S. The proposed methods in this thesis
simulate the IEC ickermeter computation beginning from the normalized volt-
age input (input to Block - 2 in Figure 2.1) to Output-5 which outputs S in
digital domain.
8
2.2 Derivation of the Mathematical Relationship between the Input
Voltage and the Flicker Sensation
The model proposed in this approach is based on the IEC icker evaluation
standard [1], whose block diagram is given in Figure 2.1. The IEC standard
[1] considers the icker-causing uctuation of the voltage as an AM modulated
signal, where the power system frequency (50 or 60Hz) is the carrier frequency
and the icker frequency is the message frequency as given in
v(t) = V sin(2ft)
_
1 +
1
2
V
V
sin(2f
f
t + )
_
(2.1)
where f is the power system (fundamental) frequency, V /V is the relative
voltage uctuation, f
f
is the icker frequency, and is the phase. An example
waveform showing the icker-causing uctuation of the voltage with V = 1,
V
V
= 0.2, f = 50Hz, and f
f
= 5Hz, is given in Figure 2.2. For the general case
of a more complicated voltage uctuation as observed in the supply voltages of
the EAF plants, the input voltage to the ickermeter can be expressed as
v(t) = V sin(2ft)
_
1 +
N
i=1
V
i
2V
sin(2f
fi
t +
i
)
_
(2.2)
where N is the number of icker frequencies, f
fi
are icker frequencies, and
i
are
the corresponding phases. The eect of the modulation appear as interharmonics
Figure 2.2: An example waveform showing the icker-causing uctuation of the
voltage
9
at beat frequencies which are ff
fi
in the spectrum since (2.2) can be rewritten
as given in
v(t) = V sin(2ft)+
N
i=1
V
i
4
{cos(2(f f
fi
)t +
i
) cos(2(f + f
fi
)t +
i
)} .
(2.3)
It is observed from (2.3) that the frequency spectrum of the voltage waveform
can be used to obtain the icker causing interharmonic frequencies and their
magnitudes. In the following paragraphs, instantaneous icker sensation com-
putation method of the IEC ickermeter will be applied to (2.2), to obtain the
relationship between the icker frequencies and the instantaneous icker sensa-
tion, S.
2.2.1 Square Demodulation and Lamp-Eye-Brain Filters
The square demodulation is applied to the signal given in (2.2) assuming the
v(t) is the normalized input waveform (input to the Block-2 in Figure 2.1)
v
2
(t) = V
2
sin
2
(2ft)
_
1 +
N
i=1
V
i
2V
sin(2f
fi
t +
i
)
_
2
(2.4)
which can be rewritten as
v
2
(t) =
1
2
V
2
(1 cos(22ft))
_
1 +
N
i=1
_
V
i
2V
_
2
sin
2
(2f
fi
t +
i
)
+2
N
i=1
N
j=i+1
(V
i
)(V
j
)
4V
2
sin(2f
fi
t +
i
)sin(2f
fj
t +
j
)
+2
N
i=1
V
i
2V
sin(2f
fi
t +
i
)
_
. (2.5)
The lamp-eye-brain chain lters are cascaded to lter the waveform given in
(2.5), whose overall response conform with band-pass lter characteristics achiev-
ing its peak magnitude at 8.8Hz. The frequencies larger than 35Hz and the dc
terms are ltered out completely [1]. Therefore, the high frequency components
which appear with the modulating eect of the cos(22ft) (f = 50 or 60Hz)
term in (2.5) will fade away at the output in addition to the dc terms. The
ltered form of (2.5) can be approximated as
10
v
2
F
(t) =
V
2
2
N
i=1
(V
i
)
2
2(2V )
2
H(2f
fi
)cos(22f
fi
t + 2
i
)
+
V
2
2
n
i=1
n
j=i+1
(V
i
)(V
j
)
4V
2
H(f
fi
f
fj
)cos(2(f
fi
f
fj
)t +
i

j
) H(f
fi
+ f
fj
)cos(2(f
fi
+ f
fj
)t +
i
+
j
)
+
V
2
2
N
i=1
V
i
V
H(f
fi
)sin(2f
fi
t +
i
). (2.6)
where H(f
fi
) is the combined frequency response of the lamp-eye-brain lters
at frequency f
fi
.
In (2.6), the terms with
(V
i
)
2
8V
2
and
(V
i
)(V
j
)
4V
2
multipliers will be negligible when
compared to the term which has the multiplication factor
V
i
V
, since
V
i
V
value
does not usually exceed 0.1 in any arc furnace voltage measured to date. Even in
the worst case of
V
i
V
= 0.5,
_
(V
i
)
2
4V
2
_
/
_
V
i
V
_
= 1/8. Therefore, the rst and the
second terms are neglected in (2.6), before proceeding to the squaring multiplier
and the rst order sliding mean lter. The squaring in the next step will also
emphasize the last term of (2.6). Therefore (2.6) is approximated as
v
2
F
(t)
N
i=1
V (V
i
)H(f
fi
)
2
sin(2f
fi
t +
i
). (2.7)
2.2.2 Squaring Multiplier and Sliding Mean Filter
The squaring multiplier output is the squared form of (2.7) which is expressed
as
v
4
F
(t) =
V
2
8
N
i=1
(V
i
)
2
H
2
(f
fi
)
V
2
8
N
i=1
(V
i
)
2
H
2
(f
fi
)cos(22f
fi
t + 2
i
)
+
V
2
4
N
i=1
N
j=i+1
((V
i
)H(f
fi
))((V
j
)H(f
fj
))cos(2(f
fi
f
fj
) +
i

j
)
V
2
4
N
i=1
N
j=i+1
((V
i
)H(f
fi
))((V
j
)H(f
fj
))cos(2(f
fi
+ f
fj
) +
i
+
j
). (2.8)
In (2.8), it is observed that at the output of the squaring multiplier of the
IEC ickermeter, there are dc components, components at twice the icker fre-
quencies and components at summations and dierences of all possible icker
frequency pairs. The sliding mean lter following the squaring multiplier is a
11
rst order low pass lter with time constant = 300ms [1]. This lter, whose
output is the instantaneous icker sensation, can be approximated as follows
[20]:
S =
1
_

v
4
F
(t)dt. (2.9)
Therefore, instantaneous icker sensation, S, will be dominated by the dc term
in (2.8) which is (
V
2
8
N
i=1
(V
i
)
2
H
2
(f
fi
)) with ripples uctuating every 300ms
and hence instantaneous icker sensation, S, can be approximated as
S
V
2
8
N
i=1
(V
i
)
2
H
2
(f
fi
). (2.10)
From (2.10) it is observed that the eect of individual icker frequencies on the
instantaneous icker sensation is approximately additive. Also phase dierences
between dierent icker frequencies have negligible eect on S. This fact has
also been pointed out based on experimental measurements by Keppler, et.al.
[21].
2.2.3 Instantaneous Flicker Sensation Determination and the Pro-
posed Method
The IEC ickermeter response for sinusoidal uctuations (frequencies and V/V
ratios) which result in unity instantaneous icker sensation (S = 1) are given
for both 230V/50Hz and 120V/60Hz systems in [1]. The proposed light icker
estimation method is applicable to both types of systems. In this research work,
230V/50Hz system is considered for eld measurements since the eld data
obtained are collected at EAF plants operating at 50Hz systems. Table 2.1
provides the ickermeter response of IEC for sinusoidal voltage uctuations in
a 230V/50Hz power system.
S is approximately the summation of the icker-causing eects of the individual
icker frequencies as given in (2.10). Contribution of each icker frequency to
S, which is represented here by S
i
, can be determined by comparing the square
of the V/V values obtained from the FFT of the voltage waveform in (2.3) to
the square of the V/V values obtained from Table 2.1 as
12
S
i
=
V
4
8
_
V
i
V
_
2
H
2
(f
fi
)
V
4
8
_
V
i
V
_
2
IEC
H
2
(f
fi
)
=
_
V
i
V
_
2
_
V
i
V
_
2
IEC
(2.11)
where
_
V
i
V
_
IEC
is obtained from Table 2.1 for the icker frequency f
fi
and
V
i
V
is obtained from the amplitudes of the FFT of the voltage waveform given in
(2.3). Frequencies around the fundamental frequency are considered (f 25Hz
to f +25Hz), where 25Hz is the maximum icker frequency given in Table 2.1.
Contributions of frequencies below f 25Hz and above f +25Hz are negligible.
Then S is obtained as the summation of all S
i
, S =
N
i=1
S
i
, where f
fi
varies
from 0.5Hz to 25.0Hz with 0.5Hz increments when the FFT window size is
100 cycles long. However, a duration of 100 cycles is too long when the non-
stationary behavior of the EAF operation is considered. The IEC Standard for
harmonics and interharmonics computation requires the FFT window sizes to be
10 cycles long [16]. Therefore, the algorithm is tested with both 100-cycle and
10-cycle long FFT-windows on simulated data and with 10-cycle long windows
Table 2.1: Normalized Flickermeter Response for Sinusoidal Voltage Fluctua-
tions, 230V/50Hz system - IEC Standard 61000-4-15
Hz V/V (%) Hz V/V (%)
0.5 2.340 10.0 0.260
1.0 1.432 10.5 0.270
1.5 1.080 11.0 0.282
2.0 0.882 11.5 0.296
2.5 0.754 12.0 0.312
3.0 0.654 13.0 0.348
3.5 0.568 14.0 0.388
4.0 0.500 15.0 0.432
4.5 0.446 16.0 0.480
5.0 0.398 17.0 0.530
5.5 0.360 18.0 0.584
6.0 0.328 19.0 0.640
6.5 0.300 20.0 0.700
7.0 0.280 21.0 0.760
7.5 0.266 22.0 0.824
8.0 0.256 23.0 0.890
8.8 0.250 24.0 0.962
9.5 0.254 25.0 1.042
13
on eld data. In case of 10-cycle windows, f
fi
varies from 5Hz to 25Hz with
5Hz increments.
S obtained from S
i
for 200ms windows overlapping 180ms or 9 cycles, which
means sampling frequency of S is 50Hz, represent the stationary behaviour of
the sliding mean lter given in (2.9). When a higher sampling rate of S is
required, the sliding mean lter can be approximated by ltering S with the
sliding mean lter as proposed by [5]. In this method, a sampling rate of 50Hz
is used for S as recommended smallest rate in [1] to reduce the computational
complexity.
2.3 Correction of the Spectral Components for Deviated Frequency
Case
DFT is a powerful and ecient algorithm to obtain the frequency distribution
of the voltage waveform. However, DFT suers from leakage eect [3, 4], when
there exist fundamental frequency deviation. In this part, initially formulation
of the leakage eect of the DFT is given, after that the proposed approach for
correction of the spectral components is explained.
2.3.1 Formulation of Leakage Eect of the Discrete Fourier Trans-
form
To formulate the spectral leakage problem of the DFT, the continuous time
signal with a single frequency component is considered:
v(t) = V
f
cos(
f
t), (2.12)
where
f
is the fundamental frequency of the power signal (
f
= 2f
f
) and
V
f
is the corresponding amplitude. Cosine is preferred here instead of sine to
make the Fourier transform equations simpler, and this does not cause any loss of
generality. Since the signals are analyzed in digital domain, this signal is sampled
with a sampling frequency f
s
, which is higher than the Nyquist frequency, and a
nite number of samples N is used for the Fourier analysis. This corresponds to
14
windowing the signal in (2.12) with a rectangular window of duration T = N/f
s
.
The windowed form of the signal is shown as:
v
w
(t) = V
f
cos(
f
t)w(t), (2.13)
where w(t) is the window function, which equals to unity for 0 < t < T, and
equals to zero elsewhere. The Fourier transform of the rectangular window
function is:
W() =
sin(T/2)
/2
e
jT/2
, (2.14)
while the Fourier transform of v(t) equals
V () = V
f
[(
f
) +( +
f
)]. (2.15)
Since the windowed function is obtained by multiplying the signal v(t) with
the window function w(t), the Fourier transform of the windowed function is
obtained by convolving their Fourier transforms as:
V
w
() = V
f
_
W( )[(
f
) + ( +
f
)]d. (2.16)
For a rectangular window, the peak appears at
f
and the zero crossing points
are at =
f
+ k2/T. If the signal is sampled at f
s
= N/T, the samples
in frequency domain correspond exactly to the points at frequencies =
f
+
k2/T according to (2.14) and only the component at
f
is non-zero in the
sampled spectrum, which is the DFT of the sampled form of v
w
(t), and hence
there is no leakage eect caused by rectangular windowing. On the other hand,
if the sampling frequency slightly diers from N/T, leakage components appear
in the spectrum due to windowing at frequencies around =
f
+ k2/T and
these components are not zero, indicating that interharmonics exist but in fact
the signal v(t) has no interharmonic components.
If the signal v(t) contains interharmonics as given in
v(t) = V
f
cos(
f
t) +V
1
cos[(
f
+
1
)t +
1
] +V
2
cos[(
f
+
2
)t +
2
], (2.17)
where V
1
and V
2
are the interharmonic amplitudes,
f
+
1
and
f
+
2
are
the interharmonic frequencies and
1
and
2
are the corresponding phases, then
15
the Fourier transform of (2.17) becomes
V () = V
f
[(
f
) + ( +
f
)]
+V
1
[(
f

1
) +( +
f
+
1
)]
+V
2
[(
f

2
) +( +
f
+
2
)].
(2.18)
The Fourier transform of the windowed signal can be obtained as:
V
w
() = V
f
_
W( )[(
f
) +( +
f
)]d
+V
1
_
W( )[(
f

1
) +( +
f
+
1
)]d
+V
2
_
W( )[(
f

2
) +( +
f
+
2
)]d
(2.19)
The rst term in (2.19) is due to the fundamental component and if the sampling
frequency equals to exactly N/T, this term will yield zero for frequency sam-
ples other than
f
. On the other hand, if the fundamental frequency deviates
slightly from 2/T, the rst term will add leakage components to the spec-
trum in addition to the interharmonic terms (2
nd
and 3
rd
terms) given in (2.19).
As observed, the leakage terms for rectangular windowing are additive and the
proposed spectral correction method is based on this fact. If the fundamental
frequency can be detected precisely, the leakage eect can be computed on syn-
thetic data and then it can be subtracted from the original signal spectrum to
correct the interharmonic values.
2.3.2 Correction of the Spectral Components
In the proposed algorithm, V
i
/V values corresponding to the icker frequencies
are obtained from the DFT of the voltage signal given in (2.3). However, DFT
suers from picket-fence eect and leakage [3, 4], when there exist interharmonic
frequencies and fundamental frequency deviation. These are commonly observed
in EAF plants due to the nature of the EAF operation. Observations on the eld
data obtained from various EAF plants [10, 22] have shown that the fundamental
system frequency tends to be lower or higher than the nominal value during
boring and melting phases and usually higher during the rening phases. From
measurements on the eld data obtained from three dierent EAF plants, it
is observed that in EAF plants, the fundamental frequency uctuates between
16
49.7Hz and 50.3Hz. In the proposed method, for the DFT computation on eld
data, window size is selected as 200ms that corresponds to the 10 cycles of the
50Hz fundamental frequency, which is compatible with the IEC Standard [16].
When the frequency deviates from the nominal value, the window size is not an
integer multiple of the actual cycle any more and this causes leakage eect on
the DFT components.
To observe the leakage eect, the algorithm is tested with simulated voltage
waveform, with sinusoidal voltage uctuation as given in (2.1) with icker fre-
quency f
f
= 10Hz, and V/V = 0.260% as it is given in Table 2.1 and funda-
mental frequency f varying from 49.5Hz to 50.5Hz. The whole data duration
is selected to be 10 minutes and P
st
is obtained from the resulting instantaneous
icker sensation, S. S is computed on every 10-cycle period with 9-cycle over-
lapping sliding windows with the proposed algorithm so that the sampling rate
of S is 50Hz as required in the IEC Standard [1]. S is expected to uctuate
around unity; however, due to the deviation in the fundamental frequency, leak-
age causes incorrect predictions of S when fundamental frequency deviates from
50Hz. The results are given in Table 2.2 under the column Without Spectral
Correction. S values given in Table 2.2 are the mean values of the S array ob-
tained for 10-minute long data. It is shown that the algorithm results in S = 1
when the fundamental frequency is exactly 50Hz. The leakage eect for fre-
Table 2.2: Eect of the System Frequency Deviation
f(Hz) Without Spectral Correction With Spectral Correction
mean(S) P
st
mean(S) P
st
49.5 17718.27 93.29 1.01 0.70
49.6 11256.69 74.22 1.01 0.70
49.7 6296.20 55.42 1.00 0.70
49.8 2787.41 36.82 1.00 0.70
49.9 695.88 18.37 1.00 0.69
50.0 1.00 0.69 1.00 0.69
50.1 695.84 18.31 1.00 0.69
50.2 2787.08 36.59 1.00 0.70
50.3 6295.04 54.90 1.00 0.70
50.4 11245.18 73.43 1.01 0.70
50.5 17713.65 92.24 1.01 0.70
17
quency deviations on S, in a smaller range from 49.9Hz to 50.1Hz with 0.01Hz
increments, when a xed 200ms-length window is used, is also illustrated in
Figure 2.3. S is found to be as high as 700, where it actually should have been
found around unity if no leakage existed.
49.9 49.92 49.94 49.96 49.98 50 50.02 50.04 50.06 50.08 50.1
0
100
200
300
400
500
600
700
Fundamental Frequency (Hz)
I
n
s
t
a
n
t
a
n
e
o
u
s

F
l
i
c
k
e
r

S
e
n
s
a
t
i
o
n
,

S
S=1
Figure 2.3: The leakage eect for frequency deviations on the instantaneous
icker sensation, S. Fundamental frequency is changed from 49.9Hz to 50.1Hz
with 0.01Hz increments. 200ms-length window is used for DFT computation.
The reason of the high error given in Table 2.2 is that the individual eects
coming from each frequency are squared and summed up as given in (2.10).
The leakage of the DFT is illustrated in Figure 2.4 for the same signal with
the same icker frequencies as given in Table 2.2. The corresponding S, whose
average values are given in Table 2.2, values are given in Figure 2.5. This gure
shows the quadratic increase in S as the fundamental frequency deviates from
the nominal value.
18
Figure 2.4: DFT leakage eect on the amplitude characteristics for the funda-
mental frequency values given in Table 2.2.
Figure 2.5: S computed out of the DFT values given in Figure 2.4 for a data of
10sec. long without spectral correction.
19
LPF +
zero-cross
detection
v[n]
FFT
N samples
Detect
Fundamental
Frequency
Obtain
Fundamental
Amplitude
Synthesize
Asin(2(pi)f
f
t)
(N samples)
FFT
N samples
n
0
(start sample)
V[k]
f
f
v
s
[n]
V
s
[k]
A
f
V[k]
zero-cross
samples
Compute Individual Conributions, S
i
,
according to IEC 61000-4-15
(from f-25Hz to f+25Hz)
Sum(S
i
)
Correct amplitudes
of V[k]
(from f-25Hz to f+25Hz
V
c
[k]
Statistical
computation
S
P
st
S
i
Figure 2.6: Block diagram of the proposed icker estimation method.
The block diagram of the proposed light icker estimation method is given in
Figure 2.6. Assume that the sampled input signal is denoted by v[n]. Since
FFT algorithm is used to realize DFT computation in this thesis, from now on
FFT will be used to explain the details of the algorithm. To go to the frequency
domain through FFT, a standard window length, N samples, is used assuming
the fundamental frequency does not deviate from the nominal value (200ms or
2s for 10-cycle or 100-cycle windows, respectively). The recommended window
length is 10 cycles for a 50Hz power system in the standard IEC-61000-4-7 [16].
N = 640 for the sampling rate of 3200Hz. The beginning of the window is
adjusted such that it starts at the nearest zero-crossing, n
0
, from negative to
positive side of the signal, so that the phase of the fundamental component is
zero. To obtain the zero-crossings a low pass lter (LPF) with a corner frequency
at 75Hz, which cleans the signal from the harmonic and interharmonic eects, is
applied rst. Then the zero-crossings are determined and FFT of the Nsample
signal is computed which is represented by V [k] in Figure 2.6. However, this
FFT contains spectral estimation errors if the fundamental frequency deviates
from the nominal value.
In order to get rid of these spectral estimation errors, a good estimation of
the fundamental frequency is necessary. The exact fundamental frequency is
computed by applying correction to the rst and the nal zero-crossings as given
20
Figure 2.7: Zero-crossing correction [23].
in [24]. This correction is illustrated in Figure 2.7 on an example zero-crossing.
The exact zero-crossing is determined by applying a linear interpolation between
the two samples as given in Figure 2.7, where Xs show the samples of the signal
and Os show the actual zero-crossing of the signal, which is computed both at the
beginning and the end of the 10-cycle period for correct fundamental frequency
computation. Then to obtain a better spectrum, a pure sinusoidal waveform of
the same frame length, N, is synthesized with the amplitude of the fundamental
A
f
and the computed fundamental frequency, f
f
, as v
s
[n] = A
f
sin(2f
f
t). As
explained in [23], the amplitude A
f
is dierent from the amplitude measured
from the DFT algorithm, if there is frequency deviation as observed in Figure 2.8.
Therefore, the correct value of A
f
at the exact fundamental frequency, f
f
, is
computed from:
2AT =
A
f
_
W(w
n
) [ w
f
+ ( + w
f
)] d
, (2.20)
Since the signals are analyzed in digital domain, signal is sampled with a sam-
pling frequency f
s
, which is higher than the Nyquist frequency in order to prevent
the aliasing eect, and a nite number of samples N is used for the Fourier
21
Figure 2.8: Amplitude of X
w
(). Xs and Os show the frequency samples with-
out and with frequency deviation, respectively [23].
analysis. This corresponds to windowing the signal with a rectangular window,
w[n] of duration T = N/f
s
. In equation (2.20), W(w) is the Fourier transform
of this window function, w[n], as given in equation (2.14), 2AT is the measured
fundamental amplitude from the FFT, w
n
is the nominal fundamental frequency
(which is 50Hz or 60Hz), A
f
is the actual amplitude of the deviated fundamen-
tal frequency. Since w
f
, 2AT, and w
n
are known, it is possible to solve for A
f
.
The FFT window is adjusted to have zero-phase at the fundamental frequency
f
f
. Therefore, the synthetic signal v
s
[n], which has zero phase, represents the
fundamental component of v[n] exactly. Then the discrete-time Fourier trans-
form of v
s
[n] is obtained, which is denoted by V
s
[k] in Figure 2.6. V
s
[k] contains
a single frequency, therefore the non-zero frequency components around the fun-
damental component show the eect of the leakage, which corresponds to the
rst term in (2.19). In the ideal case of f
f
= 50Hz, all FFT amplitudes except
the 50Hz component should be equal to zero. On the other hand, when the
fundamental frequency deviates, FFT amplitudes in the neighborhood of 50Hz
component are non-zero. Therefore, these synthetic amplitudes are used to cor-
22
rect the amplitudes of the FFT of the actual voltage waveform, V [k], around
the fundamental frequency, which are f f
fi
to f + f
fi
(from 25Hz to 75Hz
for the 50Hz case), where f is the nominal frequency. It has been shown in the
previous subsection that the leakage eect due to the fundamental frequency
shift is additive, therefore, the correction is applied so as to subtract the leakage
eect obtained from the FFT of the synthetic sine wave, V
s
[k], from the FFT
of the original voltage, V [k]. A vector subtraction is applied individually on
each of the FFT components from f f
fi
to f + f
fi
except the fundamental,
which is 50Hz or 60Hz. The corrected spectrum is denoted by V
c
[k], and the
instantaneous icker sensation, S, is computed from this corrected spectrum.
FFT amplitudes and the phases from an example voltage waveform which is
200ms long and obtained from the secondary of the EAF transformer at an EAF
plant are shown in Figure 2.9. The fundamental frequency of this waveform is
detected as 50.0336Hz. The FFT amplitudes around the fundamental frequency
of the original and the synthetic waveforms are given in Figure 2.9 (a) and (b),
respectively. The phases are given in (c) and (d), and the corrected amplitudes
are given in (e). Figure 2.10 shows an example of the vector subtraction for the
55Hz component given in Figure 2.9 to obtain the corrected FFT amplitude at
55Hz. This procedure is repeated for each frequency from f 25Hz to f +25Hz
except for the fundamental frequency.
The results with the spectral correction are given in Table 2.2 under the column
With Spectral Correction. It is observed that without spectral correction huge
errors can be observed, especially in P
st
values, when the fundamental frequency
deviates. The proposed spectral correction, on the other hand, provides unity
instantaneous icker sensation within a 1% tolerance band.
23
Figure 2.9: Spectral amplitude correction.
Figure 2.10: Amplitude correction of the 55Hz component in Figure 2.9. Cor-
rection is achieved by vector subtraction of the synthetic component from the
original component using both phases and amplitudes given in Figure 2.9.
24
To demonstrate the usefulness of the proposed spectral correction method, an
example instantaneous icker sensation computation is given in Figure 2.11 with
and without spectral correction applied on the 10-minute duration voltage ob-
tained at the MV side of an AC EAF plant. The frequency deviation is also given
in Figure 2.11 (a) to illustrate the increase in the error due to the FFT leakage
when the frequency deviates more from the 50Hz value between minutes 3.5 and
6. Figure 2.11 (b) shows S computed without the spectral correction compared
with the S obtained from the digital realization of the IEC ickermeter. In part
(c), S computed with spectral correction is given. It is observed that spectral
correction helps to estimate S very close to the IEC result especially when the
frequency deviation is signicant.
Figure 2.11: Eect of frequency variation on estimation of S. (a) Frequency
variation, (b) estimation of S without spectral correction compared with the
IEC ickermeter result, (c) estimation of S with the proposed method (with
spectral correction) compared with the IEC ickermeter result.
25
2.4 Verication of the Proposed Method
For verication of the proposed method, a digital realization of the IEC icker-
meter given in [9] is used to compare the response of the proposed method to
the actual IEC ickermeter behavior. Both simulated and eld data are used in
order to test the proposed method.
2.4.1 Verication with Simulated Data
Simulated data is produced by modulating the sine wave at 50Hz system fre-
quency as given in (2.1) with various icker frequencies and the corresponding
V/V values in Table 2.1 with a sampling frequency of 3200Hz. Data length
is selected as 10 minutes. S is computed with the proposed light icker esti-
mation method and the digital IEC ickermeter realization and the results are
compared. Three tests are done to test the proposed algorithm with simulated
data.
2.4.1.1 Test 1: Single Flicker Frequency Case
A single sinusoidal icker frequency case is given in (2.1). It is expected that S
should be close to unity since Table 2.1 given in the IEC standard [1] is for unity
S response to sinusoidal voltage uctuations. The experiment is repeated with
FFT window sizes of both 200ms and 2s (corresponding to 10 cycles and 100
cycles of the ideal system frequency), resulting in an FFT frequency resolution
of 5Hz and 0.5Hz respectively. In fact, 2s is a long time for the actual EAF
voltage to be stationary and the recommended window size is 200ms according
to [16]; however, in case of simulated signals, data can be kept stationary as long
as required, so both resolutions are used to test the algorithm. In case of 0.5Hz
resolution, summation of all S
i
given in (2.11) from 0.5Hz to 25Hz are used to
obtain S, whereas V/V values corresponding to only 5, 10, 15, 20 and 25Hz are
added up to obtain S in case of 5Hz resolution. Comparison of the results with
those of the IEC ickermeter are given in Table 2.3 for both resolution values.
26
Table 2.3: Proposed light icker estimation algorithm response for sinusoidal
voltage uctuations
Input voltage uctuation to obtain Proposed ickermeter response
S = 1 with IEC ickermeter 5Hz resolution 0.5Hz resolution
icker freq.(Hz) V/V % S error(%) S error(%)
5.0 0.40 1.00 0.0 1.00 0.0
10.0 0.26 1.00 0.0 1.00 0.0
12.0 0.31 1.01 1.0 1.00 0.0
15.0 0.43 1.00 0.0 1.00 0.0
20.0 0.70 1.00 0.0 1.00 0.0
25.0 1.04 1.00 0.0 1.00 0.0
2.4.1.2 Test 2: Rectangular Voltage Fluctuations-Combinations of
Flicker Frequencies
Rectangular uctuation of the voltage is given as
v(t) = V sin(2ft)
_
1 +
V
2V
signum[sin(2f
f
t)]
_
. (2.21)
Rectangular uctuation can be considered as the linear combination of odd
harmonic sinusoidal uctuations due the Fourier Series expansion of the rect-
angular periodic waveform. This corresponds to representing the rectangular
uctuations as given in (2.2). Normalized ickermeter response for rectangu-
lar voltage uctuations is also tabulated in [1]. To test the proposed method,
simulated waveform in (2.21) is obtained with V/V values taken from the
Table Normalized ickermeter response for rectangular uctuations for 230V
lamp and 50Hz systems in the IEC Standard [1]. This table is provided in Ap-
pendix A, Table A.1. It is observed that the proposed method estimates S with
a maximum error of 6% as given in Table 2.4.
2.4.1.3 Test 3: Homogeneity of the Flickermeter Response
The Standard IEC-61000-4-15 requires that the ickermeter response should
have the homogeneity property, i.e., whenever the V/V value is doubled, the
resulting icker severity P
st
should also be doubled [1]. Homogeneity of the pro-
posed algorithm is checked by applying twice as much and half of the V/V
27
Table 2.4: Proposed icker estimation algorithm response for rectangular voltage
uctuations
5.0 0.29 0.98 -2.0 0.98 -2.0
10.0 0.21 1.03 3.0 1.04 4.0
12.0 0.25 1.02 2.0 1.02 2.0
15.0 0.34 1.04 4.0 1.06 6.0
20.0 0.55 0.99 -1.0 1.02 2.0
values with the corresponding icker frequencies given in Table 2.1. If the homo-
geneity of the proposed algorithm holds, it is expected that P
st
value is doubled
and halved, when V/V is doubled and halved, respectively. It is observed
that the proposed method estimates P
st
values with minor error as given in Ta-
ble 2.5. The error in this table is dened as the percentage dierence between
twice/half of the P
st
computed by the IEC ickermeter and the P
st
obtained by
the proposed method when V/V is doubled/halved.
2.4.2 Verication on AC Arc Furnace Data
To verify the eectiveness of the proposed algorithm in estimating the light
icker, experiments on eld data obtained at three dierent EAF plants are
used. The proposed light icker estimation algorithm is used with 5Hz frequency
resolution (200ms windows). The data is obtained during eld measurements
carried out through the National Power Quality Project of Turkey [22, 26]. EAF
voltages are collected at a sampling rate of 3200Hz [25]. EAF rms voltages
obtained at the MV side of the EAF transformer at three plants are given with
the instantaneous icker sensation, S, values obtained from both the digital
implementation of the IEC ickermeter and the proposed light icker estimation
algorithm in Figures 2.12, 2.13, and 2.14. In these gures, (a) shows the rms
voltages with the P
st
computed intervals, (b) shows the frequency variation, and
(c) shows S computed with both the proposed method and the IEC ickermeter
for comparison. It is observed that the proposed method gives good estimates
28
Table 2.5: Homogeneity of the proposed light icker estimation algorithm
Flicker frequency (Hz) V/V % %
sinusoidal uctuation 230V/50Hz system P
st
Error
2.340 0.98
0.5 2 2.340 1.96 0.00
0.5 2.340 0.49 0.00
0.754 0.59
2.5 2 0.754 1.17 +0.84
0.5 0.754 0.30 -1.67
0.398 0.69
5.0 2 0.398 1.38 0.00
0.5 0.398 0.35 -1.45
0.260 0.69
10.0 2 0.260 1.38 0.00
0.5 0.260 0.35 -1.45
0.312 0.73
12 2 0.312 1.46 0.00
0.5 0.312 0.37 0.00
0.432 0.69
15 2 0.432 1.38 0.00
0.5 0.432 0.35 -1.45
0.584 0.78
18 2 0.584 1.55 +0.64
0.5 0.584 0.39 0.00
0.700 0.69
20 2 0.700 1.38 0.00
0.5 0.700 0.35 -1.45
1.042 0.69
25 2 1.042 1.38 0.00
0.5 1.042 0.35 -1.45
of the instantaneous icker sensation, S. More detailed views of S for the rst
two intervals for Figures 2.12, 2.13, and 2.14 are given in Figures 2.15, 2.16, and
2.17, respectively for a better comparison. To give a quantitative comparison
of the algorithm, short term icker severity, P
st
, is computed for the intervals
given in the gures and the results are compared with the results of the digital
realization of the IEC ickermeter as given in Tables 2.6, 2.7 and 2.8. The
results indicate good agreement between the IEC ickermeter and the proposed
method.
29
Figure 2.12: Plant-1 (a) rms voltage at the MV side of the EAF Plant with the
P
st
computed intervals, (b) fundamental frequency variation, (c) instantaneous
icker sensation, S.
Table 2.6: Short term icker severity, P
st
, computed inside the intervals given
in Fig 2.12 with the proposed method and the digital realization of the IEC
ickermeter
INTERVAL 1 INTERVAL 2 INTERVAL 3 INTERVAL 4 INTERVAL 5
IEC Flickermeter 15.4958 9.5497 13.2904 2.6938 1.8163
Proposed Method 15.4131 9.6166 13.5436 2.6922 1.9905
Error % +0.53 -0.70 -1.91 +0.06 -9.59
30
P
st
icker sensation, S.
st
in Figure 2.13 with the proposed method and the digital realization of the IEC
ickermeter
Proposed Method 23.3268 22.5856 11.9070 9.4001 10.9129
Error % -4.51 -5.40 -2.78 -1.34 -2.61
31
P
st
icker sensation, S.
st
in Figure 2.14 with the proposed method and the digital realization of the IEC
ickermeter
Proposed Method 11.7285 15.6946 9.2464 8.8416 7.4208
Error % -0.74 +2.14 -2.73 -6.83 -14.49
32
Figure 2.15: Plant-1 instantaneous icker sensation, S, for Interval-1 and
Interval-2 in Figure 2.12 for a more detailed comparison of the proposed method
and the IEC ickermeter.
33
The calculation procedures are analyzed on a personal computer (PentiumIV,
2.20GHz). All algorithms are implemented using MATLAB running on Win-
dows XP. The total computation time for a 10min input data is 3.5s. This shows
the computational eciency of the method when compared with the computa-
tion times given in [2], which are inside the range from 7.96s to 0.91s for an input
data length of 4s with MATLAB running on a PentiumIV, 2.20GHz personal
computer.
The accuracy obtained on eld data are comparable with the results given in
[5] and in [2]. The mean error is 2.56%, 3.33%, and 5.39% for Plant-1, Plant-2
and Plant-3, respectively. In [2], mean error is around 6% and it varies between
0.44% and 5.94% in [5], which shows that the proposed method gives satisfactory
results.
For a more detailed comparison, also the method proposed in [5] is compared
with the proposed method in this research work. Initially the method in [5]
is implemented and the data given in Figure 2.12 are used to compare the P
st
34
results of the two methods. Table 2.9 presents the results. It is shown that the
proposed method provides comparable results with the methods presented in
the literature, which shows that the method is successful in estimating the light
icker.
Table 2.9: Comparison of the short term icker severity, P
st
, computed inside
the intervals given in Fig 2.12 with the proposed method, the method given in
[5] and the digital realization of the IEC ickermeter
Error %
(Method in [5]) +1.81 -3.72 -1.92 +0.08 -8.86
Error %
(Proposed Method) +0.53 -0.70 -1.91 +0.06 -9.59
35
CHAPTER 3
A KALMAN FILTERING BASED APPROACH FOR
LIGHT FLICKER EVALUATION OF POWER
SYSTEMS
Flicker evaluation method based on interharmonic spectrum is provided in detail
in Chapter 2 and in [7]. It has been shown that the eect of each interharmonic
frequency, f
fi
, to the instantaneous icker sensation S is
S
i
=
V
4
8
_
V
i
V
_
2
H
2
(f
fi
)
V
4
8
_
V
i
V
_
2
IEC
H
2
(f
fi
)
=
_
V
i
V
_
2
_
V
i
V
_
2
IEC
(3.1)
and S is obtained as the summation of all S
i
, S =
N
i=1
S
i
, where f
fi
varies
from 5Hz to 25Hz with 5Hz increments for 10-cycle windows.
In the method presented in this chapter, the frequency domain components of
the voltage waveform are obtained by Kalman ltering and these components
are used to obtain the light icker using the relationship given in (3.1). In order
to describe the usage of the Kalman lters in the proposed method, the general
characteristic of Kalman lters is explained in Section 3.1 based on [27].
3.1 General Characteristics of Kalman Filters
The Kalman lter is a set of mathematical equations. It provides an ecient
recursive algorithm to estimate the state of a process, by minimizing the mean
36
of the squared error. The Kalman lter supports estimations of past, present,
and even future states. It can achieve success even when the precise nature of
the modeled system is unknown.
As it is clearly explained in [27], the aim of the Kalman lter is to estimate the
state variable x R
n
of a discrete time controlled process which is shown by
the linear stochastic dierence equation
x
k
= Ax
k1
+Bu
k1
+ w
k1
(3.2)
with a measurement z R
m
which is shown as
z
k
= Hx
k
+ v
k
(3.3)
The random variable w
k
represents the process noise and the random variable
v
k
represents the measurement noise which are assumed to be independent from
each other, white and with normal probability distributions. The process noise
covariance is represented by Q, and the measurement noise covariance is repre-
sented by R.
The aim of the n x n matrix A in (3.2) is to relate the state at the previous time
step k 1 to the state at the current step k. The aim of the n x l matrix B in
(3.2) is to relate the optional control input u R
l
to the state x. In (3.3), the
m x n matrix H relates the state to the measurement z
k
.
Initially a priori state estimate at step k, x
k
R
n
, and a posteriori state estimate
at step k, x
k
R
n
, are dened with given measurement z
k
. Then a priori and
posteriori estimate errors are dened as
e
k
= x
k
x
k
(3.4)
and
e
k
= x
k
x
k
(3.5)
P
k
= E[e
k
e
k
T
] is then called as a priori estimate error covariance and P
k
=
E[e
k
e
k
T
] is called as a posteriori estimate error covariance.
At each iteration the estimate of the state variable is updated with measurement
z
k
as given below.
x
k
= x
k
+ K(z
k
H x
k
) (3.6)
37
In (3.6), H x
k
is the predicted measurement and z
k
is the actual measurement.
The dierence (z
k
H x
k
) is called the residual which shows the discrepancy
between the predicted and the actual measurements. In this equation, the n x
m matrix K is the gain, which minimizes the posteriori error covariance, P
k
.
One form of the resulting K, which minimizes P
k
is given as
K
k
= P
k
H
T
(HP
k
H
T
+ R)
1
. (3.7)
Using these new parameters, the linear Kalman lter is implemented as ex-
plained below.
The Kalman lter estimates a process by using a form of feedback control.
Therefore, the equations for the Kalman lter fall into two groups. First group
consists of time update equations and second group consists of measurement
update equations. The time update equations are responsible for projecting for-
ward the current state and error covariance estimates in order to obtain a priori
estimates for the next time step [27]. On the other hand, the measurement up-
date equations are responsible for the feedback part. This part incorporates a
new measurement into the a priori estimate to obtain an improved a posteriori
estimate [27]. Therefore, it is possible to declare that time update equations are
responsible for prediction whereas measurement update equations are responsi-
ble for the correction.
The linear Kalman lter time update equations are given as follows :
x
k
= A x
k1
+ Bu
k1
P
k
= AP
k1
A
T
+ Q
while the linear Kalman lter measurement update equations are given as below:
K
k
= P
k
H
T
(HP
k
H
T
+R)
1
x
k
= x
k
+ K
k
(z
k
H x
k
)
P
k
= (I K
k
H)P
k
The complete operation of the linear Kalman lter is given in Figure 3.1. Each
step of this gure is explained in [27] clearly.
38
Figure 3.1: Block diagram of the operation of the Kalman Filter [27].
The extended Kalman lter is a little bit dierent from the linear Kalman lter
algorithm. When the process to be estimated and or the measurement relation-
ship to the process is nonlinear, extended Kalman lter is used. Thus, a Kalman
lter that linearizes about the current mean and covariance is referred to as an
EKF [27].
The extended Kalman lter time update equations are given as follows :
x
k
= f( x
k1
, u
k1
, 0)
P
k
= A
k
P
k1
A
k
T
+ W
k
Q
k1
W
k
T
while the extended Kalman lter measurement update equations are given as
below :
K
k
= P
k
H
k
T
(H
k
P
k
Hk
T
+ V
k
R
k
V
k
T
)
1
x
k
= x
k
+ K
k
(z
k
h( x
k
, 0))
P
k
= (I K
k
H
k
)P
k
Figure 3.2 shows the complete operation of the extended Kalman lter. The
detailed explanation of the steps of the EKF algorithm is included in [27].
39
Figure 3.2: Block diagram of the operation of the Extended Kalman Filter [27].
In the proposed method, an EKF is used to estimate the voltage envelope and
then a LKF is used to obtain the spectral amplitudes out of the voltage envelope
causing light icker.
3.2 Obtaining the Spectral Decomposition based on Kalman Filter-
ing
The block diagram of the proposed spectral decomposition algorithm, which is
also explained in [18], is given in Figure 3.3. To obtain spectral decomposition
by Kalman ltering, a standard window length is used. When the fundamental
frequency deviates from the nominal value, the window size is not an integer
multiple of the actual cycle any more. In such a case, when DFT algorithm
is used to obtain the spectral decomposition, even in a very small deviation of
the fundamental frequency case, leakage eect occurs on the DFT components,
which causes signicant error as explained in detail in Section 2.3. However, in
the proposed spectral decomposition method, fundamental frequency deviation
is not so critical in terms of the estimated spectral amplitudes. Since Kalman
lters are used to obtain the spectral decomposition, no leakage eect problem
occurs.
40
A(t)cos(2f
o
t+
o
)
A(t)
f
o
(t)
Spectral amplitudes
(from 0Hz to 45Hz)
Figure 3.3: Block diagram of the proposed spectral decomposition method.
In the proposed method, window length is selected as 200ms, which is the
recommended window length in the harmonic and interharmonic computation
standard of IEC, IEC-61000-4-7 [16]. Then for each time window of 200ms,
the spectral decomposition is obtained by applying two types of Kalman l-
ters, EKF and LKF, consecutively [18, 19]. The EKF and LKF algorithms are
implemented based on [27]. The voltage uctuation containing interharmonic
frequencies around the fundamental frequency can be modeled as given in [1]
v(t) = cos(2f
o
t +
o
)
_
V +
N
i=1
V
i
2
cos(2f
i
t +
i
)
_
. (3.8)
The voltage signal in (3.8) represents the voltage uctuation especially observed
in EAF plants, which contains several interharmonic frequencies. Figure 3.4
shows an example voltage waveform containing several interharmonic frequen-
cies. As it is clearly seen from this gure that the signal model in (3.8) can also
be modeled as a sinusoidal voltage waveform with time varying amplitude A(t)
and it is shown as
v(t) = A(t)cos(2f
o
t +
o
),
= A(t)cos(
o
)cos(2f
o
t) A(t)sin(
o
)sin(2f
o
t). (3.9)
In equation (3.9), A(t) is the envelope of the input signal which can also be
dened as the time varying amplitude of the input signal. To obtain the spec-
tral decomposition of the input signal, v(t), rst the voltage envelope, A(t) is
41
Figure 3.4: Voltage waveform containing dierent interharmonic frequencies.
obtained by applying an EKF on the waveform [18]. A(t) is assumed to contain
N frequency components, which cause interharmonics around the fundamental.
In the second step of the proposed algorithm, these frequency components are
determined by applying a LKF as shown in the block diagram of the proposed
method, Figure 3.3.
Kalman lters are used to estimate the state x R
n
of a discrete time controlled
process which is governed by a linear stochastic dierence equation as shown as
x
k
= Ax
k1
+Bu
k1
+ w
k1
(3.10)
with a measurement z R
m
, which is shown as
z
k
= Hx
k
+ v
k
. (3.11)
The random variable w
k
represents the process noise and the random variable v
k
represents the measurement noise, which are assumed to be independent from
each other, white and with normal probability distributions. However, when the
42
process to be estimated or when the measurement relationship to the process is
nonlinear, extended Kalman lters are used as mentioned in the previous Section
3.1. In the proposed method, EKF is used to obtain the envelope, A(t), therefore
the state variables for this EKF algorithm are determined as given below :
x
1
: in-phase component of v(t), (A(t)cos(
o
)),
x
2
: quadrature-phase component of v(t), (A(t)sin(
o
)),
x
3
: fundamental frequency of v(t), (f
o
).
The EKF is applied on the voltage waveform for N sample windows as shown in
Figure 3.3. Since in the proposed method a standard window length of 200ms is
used as recommended in the standard IEC-61000-4-7 [16], each window consists
of 0.2f
s
samples, where f
s
is the number of samples per second. For verication
of the proposed method, all tests are achieved with a sampling frequency of
3.2kHz, which means that EKF is applied on the windows of length 640 samples.
For each time step, the voltage phasor, which is represented by x
1
+jx
2
, is rotated
by the amount of 2x
3
t. The state transition of x
1
and x
2
can be written as ;
x
1
(k + 1) + jx
2
(k + 1) = [x
1
(k) +jx
2
(k)]e
j2x
3
(k)t
. (3.12)
According to the equation (3.12), the nonlinear state transition equation is ob-
tained as
_
_
x
1
(k + 1)
x
2
(k + 1)
x
3
(k + 1)
_
_
=
_
_
x
1
(k)cos(2x
3
(k)t) x
2
(k)sin(2x
3
(k)t)
x
1
(k)sin(2x
3
(k)t) x
2
(k)cos(2x
3
(k)t)
x
3
(k)
_
_
. (3.13)
Since the fundamental frequency observed in the power systems is not usually
constant in practice, it is also considered as the state variable, x
3
, to be esti-
mated in the proposed method. The results show that the proposed method
is successful in estimating the fundamental frequency of the voltage waveform.
The advantage of the proposed method is that there is no fundamental frequency
computation requirement, but it is only determined as a state variable in the
EKF algorithm.
43
The EKF equations in the proposed method are implemented according to [27].
Since the voltage signal is a nonlinear function of the frequency, the state transi-
tion and the measurement equations are also nonlinear functions of x
3
. In EKF
algorithm state transition and measurement equations are combined together in
accordance with the EKF model explained in [28]. The EKF model is realized,
and by applying EKF on the voltage waveform, its envelope, A(t), is obtained.
The linearized state transition matrix and the linearized measurement vector
are given as,
=
_
_
cos(2 x
3
t) sin(2 x
3
t) 2t( x
1
sin(2 x
3
t) x
2
cos(2 x
3
t))
sin(2 x
3
t) cos(2 x
3
t) 2t( x
1
cos(2 x
3
t) x
2
sin(2 x
3
t))
0 0 1
_
_
, (3.14)
and
H =
_
cos(2 x
3
t) sin(2 x
3
t) 2t( x
1
sin(2 x
3
t) + x
2
cos(2 x
3
t))
_
, (3.15)
respectively, where x
i
is a priori state estimate of the i
th
state.
The EKF algorithm is applied by using the linearized state transition matrix, ,
and the linearized measurement vector, H. At the output of the EKF, the state
variables of the in-phase component of the voltage envelope, the quadrature-
phase component of the voltage envelope, and the state variable showing the
fundamental frequency deviation are obtained for each window of 200ms. Using
the estimated state variables of the state vector, which is the output of the EKF
algorithm, the envelope of the voltage waveform and the fundamental frequency
variation are obtained. The envelope of the voltage waveform is formed by using
the state variables x
1
and x
2
as
A(t) =
_
x
1
(t)
2
+ x
2
(t)
2
. (3.16)
The variation of the state variable x
3
gives the time-varying fundamental fre-
quency, which is represented as
f(t) = x
3
(t). (3.17)
44
The output of the EKF algorithm gives the fundamental frequency variation
and the envelope of the voltage waveform. After the EKF, comes a second
Kalman lter, which is a linear Kalman lter (LKF). Spectral decomposition of
the voltage waveform is obtained by using this LKF. The input of the LKF is the
envelope of the voltage waveform, A(t), which is the output of EKF and contains
all the frequency components causing interharmonics around the fundamental.
The mathematical representation of the envelope of the voltage waveform, A(t),
is
A(t) =
_
V +
N
i=1
V
i
2
cos(2f
i
t +
i
)
_
. (3.18)
A(t) is the measurement (z
k
in the measurement equation (3.11)) of the LKF
algorithm. Spectral decomposition of A(t) which is obtained by using a LKF
gives the spectral amplitudes at frequencies f
i
, directly. Since the aim is to
obtain the spectral amplitudes at the interharmonics around the fundamental,
the corresponding frequency region for the frequencies, f
i
, is between 0 45Hz.
In equation (3.18), constant V is considered as a DC voltage signal. Therefore,
A(t) is considered to be represented as the summation of ten sinusoidal voltage
uctuations (0, 5, 10, 15, , 45Hz), and hence the results are obtained with
5Hz resolution in the proposed method. Thereby, the results can be compared
with 5Hz resolution results of the method which is explained in [23]. Higher
frequency resolution can be preferred for more precise estimates, although fre-
quency resolution increase has a cost of increasing the matrix dimensions of the
state equations. For one of these sinusoidal voltage uctuations, the in-phase
and the quadrature-phase components, which are determined as the two state
variables for the corresponding voltage uctuation in the LKF model, are ob-
tained from sinusoidal expansion of the voltage uctuation as
V
i
2
cos(2f
i
t +
i
) =
V
i
2
cos(
i
)cos(2f
i
t)
V
i
2
sin(
i
)sin(2f
i
t). (3.19)
In (3.19), states are dened as follows:
x
1
= in-phase component of the sinusoidal uctuation (
V
i
2
cos(
i
)),
x
2
= quadrature-phase component of the sinusoidal uctuation (
V
i
2
sin(
i
)),
i = 1, 2, , 10 represents the frequencies 0, 5, 10, 15, , 45 Hz.
45
The state vector of the LKF, x
k
, is dened as
x
k
=
_
x
1,1
x
2,1
x
1,2
x
2,2
x
1,10
x
2,10
_
T
k
(3.20)
where superscript T denes transpose of the vector given in brackets and,
x
1,i
= in-phase component of the sinusoidal uctuation at frequency, f
i
x
2,i
= quadrature-phase component of the sinusoidal uctuation at frequency,
f
i
(f
i
= {0, 5, 10, 15, 20, , 45}).
Then the state transition equation for the twenty-state Kalman lter model is
given as
x
k+1
=
_
_
1 0 0
0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1
_
_
x
k
+ w
k
. (3.21)
Since the state variables x
1,i
and x
2,i
represent the in-phase and quadrature-
phase components of the voltage phasor at frequency f
i
, the measurement equa-
tion of the LKF is formed as
z
k
= x
k
+ v
k
, (3.22)
where the corresponding measurement vector, , is
=
cos(2f
1
t) sin(2f
1
t) cos(2f
2
t) sin(2f
2
t) cos(2f
10
t) sin(2f
10
t)
. (3.23)
In this twenty-state LKF model, which is formed by equations (3.21) and (3.22),
w
k
represents the process noise vector and v
k
represents the measurement noise
vector. The spectral amplitude at frequency f
i
is obtained by using the corre-
sponding state variables x
1,i
and x
2,i
, which are calculated for each time window.
The spectral amplitude at frequency f
i
for the corresponding time window is de-
termined by,
V
i
2
=
_
x
2
1,i
+ x
2
2,i
. (3.24)
46
3.2.1 Verication of Kalman Filtering Method for Interharmonic De-
tection
The proposed spectral decomposition method is veried by computations on
both simulative data and eld data. For eld data verication, the interhar-
monic content obtained by the proposed algorithm, the method explained in
[23], and the standard FFT algorithm are compared with the interharmonic
content obtained from FFT when the FFT window length is adjusted to t an
integer multiple of the fundamental frequency as given in [29]. All algorithms
of the proposed method are implemented using MATLAB running on Windows
XP. The total computation time for a 1 min input data is almost 75 s. However,
this computation time can be reduced much more by using a faster compiling
environment such as C and a faster processor. The code can also be optimized
for ecient computation.
3.2.1.1 Analysis on Simulative Data
Two dierent synthetic power signals with various interharmonic contents are
considered in this part. The error between the correct rst interharmonic value,
both computed as group interharmonic and subgroup interharmonic according to
the IEC standard [16], are compared with the results of the proposed approach,
the method explained in [23] and the standard FFT approach with the window
length of 200ms. First interharmonics are considered, because the leakage eect
due to the fundamental frequency deviation mainly aect the interharmonics
around the fundamental frequency. The nominal value of the fundamental fre-
quency is 50Hz and it is assumed to be varying from 49.8Hz to 50.2Hz since
this has been observed in the eld voltage data collected at the arc furnace
plants [7, 10]. Extreme cases of 49.5Hz and 50.5Hz are also considered in the
simulative analysis. In the IEC standard [16], for 50Hz power system, group
interharmonic computation is given as
C
2
ig,n
=
9
i=1
C
2
k+i
, (3.25)
47
and the subgroup interharmonic computation is given as
C
2
isg,n
=
8
i=2
C
2
k+i
, (3.26)
where k represents the spectral component index for 5Hz resolution (i.e. k = 10
for the fundamental of 50Hz), n is the harmonic number (i.e. n = 1 for the
fundamental), C
k+i
is the rms value of the (k+i)
th
spectral component, and C
ig,n
and C
isg,n
are the rms values of the interharmonic centered group and subgroup
of order n, respectively (for example, the subgroup between n = 5 and n = 6 is
designated as C
isg,5
). The tests in this part are achieved for n = 1 case.
Case-1: Single interharmonic case
The time window of 200ms results in 5Hz frequency resolution for DFT compo-
nents. Therefore, to obtain the interharmonic frequencies around the fundamen-
tal frequency using Kalman ltering, twenty-state Kalman lter model is used
(from f = 0 to 45Hz with 5Hz resolution both for in-phase and quadrature-
phase components). In this case the signal x(t) = sin(2ft) + 0.5sin(265t) is
considered with the fundamental frequency f. If no leakage eect existed, i.e.
f = 50.00Hz, the interharmonic subgroup-1 (interharmonics between the fun-
damental frequency and the 2
nd
harmonic) and the interharmonic group-1 would
both be equal to 0.50 when the FFT algorithm is applied, since the amplitude
of the only interharmonic frequency is 0.50. Otherwise, when the fundamen-
tal frequency is not exactly 50Hz, leakage eect is observed. Subgroup and
group interharmonic computations are achieved according to [16]. The results
are given in Table 3.1. It is observed in Table 3.1 that the method in [23] has
no error for both group and subgroup interharmonic computations, which is
due to the fact that this method provides the exact interharmonic content of
the signal when the fundamental frequency is determined precisely. With the
classical FFT approach with a standard window of 200ms long, the error in the
rst subgroup interharmonic value may reach 3.12%, when the frequency devi-
ates between 49.8Hz and 50.2Hz, which was the case observed in the eld data
collected at the HV transformers supplying the arc furnace plants [10]. For the
extreme case of 49.5Hz, standard FFT results in an error of 7.76% and 9.40% for
subgroup and group rst interharmonics, respectively. Finally for the Kalman
48
Table 3.1: Both group and subgroup interharmonic-1 computed with the stan-
dard FFT computation (FFT), the method in [23] and the proposed method
(PM) for x(t) = sin(2ft) + 0.5sin(265t)
f(Hz) Interharmonic-1 subgroup Interharmonic-1 group
FFT error (%) Mtd in [23] error (%) PM error (%) FFT error (%) Mtd in [23] error (%) PM error (%)
49.50 0.54 7.76 0.50 0.00 0.4966 0.68 0.55 9.40 0.50 0.00 0.4980 0.40
49.80 0.52 3.12 0.50 0.00 0.4974 0.52 0.52 3.44 0.50 0.00 0.4983 0.34
49.90 0.51 1.54 0.50 0.00 0.4976 0.48 0.51 1.62 0.50 0.00 0.4984 0.32
49.95 0.51 0.76 0.50 0.00 0.4978 0.44 0.50 0.78 0.50 0.00 0.4985 0.30
50.00 0.50 0.00 0.50 0.00 0.4979 0.42 0.50 0.00 0.50 0.00 0.4986 0.28
50.05 0.49 0.74 0.50 0.00 0.4979 0.42 0.48 0.72 0.50 0.00 0.4986 0.28
50.10 0.49 1.46 0.50 0.00 0.4979 0.42 0.49 1.38 0.50 0.00 0.4986 0.28
50.20 0.48 2.84 0.50 0.00 0.4977 0.46 0.49 2.44 0.50 0.00 0.4985 0.30
50.50 0.47 6.10 0.50 0.00 0.4963 0.74 0.48 3.40 0.50 0.00 0.4977 0.46
ltering based method, the error value is around 0.45% and 0.30% for sub-
group and group interharmonic computations respectively as given in Table 3.1,
which shows the Kalman ltering based method is successful in estimating the
interharmonics without applying two FFT algorithms and precise fundamental
frequency detection as in the case of [23]. The maximum errors are observed
for f = 50.5Hz and they are 0.74% and 0.46% for the rst subgroup and group
interharmonics, respectively, as shown in Table 3.1. Only the rst two decimal
digits are shown in all tables for the sake of simplicity.
Case-2: Interharmonics at various frequencies case
In this case the signal x(t) = sin(2ft) + 0.7sin(253t) + 0.6sin(255t) +
0.5sin(258t) +0.4sin(262t) +0.3sin(265t) is considered with the fundamen-
tal frequency f. If no leakage eect existed, the interharmonic subgroup-1 would
be equal to 0.62 and the interharmonic group-1 would be 0.96 (these are com-
puted for the case of f = 50.00Hz). The aim of this case is to test the proposed
method with a signal similar to the eld data. There are various interharmonics
between the fundamental and the second harmonics with decreasing amplitudes
from the fundamental frequency towards the second harmonic.
49
Table 3.2: Both group and subgroup interharmonic-1 computed with the stan-
dard FFT computation (FFT), the method in [23] and the proposed method
(PM) for x(t) = sin(2ft) + 0.7sin(253t) + 0.6sin(255t) + 0.5sin(258t) +
0.4sin(262t) + 0.3sin(265t)
f(Hz) Interharmonic-1 subgroup Interharmonic-1 group
FFT error (%) Mtd in [23] error (%) PM error (%) FFT error (%) Mtd in [23] error (%) PM error (%)
49.50 0.68 9.01 0.62 0.00 0.5682 4.12 1.07 11.59 0.96 0.00 0.9271 3.92
49.80 0.64 3.28 0.62 0.00 0.6948 0.96 1.01 4.55 0.96 0.00 0.9450 3.29
49.90 0.63 1.54 0.62 0.00 0.6052 0.61 0.99 2.26 0.96 0.00 0.9504 3.20
49.95 0.63 0.76 0.62 0.00 0.6107 0.54 0.98 1.68 0.96 0.00 0.9532 1.45
50.00 0.62 0.00 0.62 0.00 0.6163 0.33 0.96 0.00 0.96 0.00 0.9558 2.98
50.05 0.62 0.71 0.62 0.00 0.6219 0.56 0.95 1.11 0.96 0.00 0.9582 3.00
50.10 0.61 1.37 0.62 0.00 0.6276 0.74 0.94 2.19 0.96 0.00 0.9603 3.03
50.20 0.61 2.47 0.62 0.00 0.6387 1.66 0.92 4.25 0.96 0.00 0.9633 3.09
50.50 0.60 4.08 0.62 0.00 0.6637 4.00 0.88 9.21 0.96 0.00 0.9575 3.14
It is observed in Table 3.2 that the proposed method is again successful in
computing the rst interharmonic value. The maximum error observed in the
fundamental frequency range of 49.5Hz and 50.5Hz is 4.12% and 3.92% for
subgroup and group rst interharmonics, respectively, as shown in Table 3.2.
3.2.1.2 Analysis on Field Data
Field data are collected at an arc furnace plant at the MV side. Sampling rate is
3.2kHz (i.e. 640samples/10cycles for the ideal case of f = 50Hz) as in the case
in [23]. The proposed algorithm is applied to the voltage waveform collected
during the boring phase of the arc furnace operation, where the frequency de-
viates signicantly. The rms voltage and the corresponding frequency at one
phase is shown in Figure 3.5. The corresponding frequency of one phase is com-
puted using Kalman ltering. The fundamental frequency is determined as the
state variable of the EKF and the fundamental frequency deviation is obtained
at the output of the EKF. During this test process, the same approach given in
[23] is followed. The voltage is normalized to have unity amplitude at the 50Hz
component to make the results comparable with the analysis on simulative data
case.
50
Figure 3.5: Magnitude and the fundamental frequency variation of the eld data
collected at the MV side of the power system supplying an arc furnace plant.
Fundamental frequency deviation is obtained by using EKF.
To form a reference for the interharmonic computation, resampling through
interpolation is achieved on the 10-cycle windows of the waveform as proposed
in [29]. The data is resampled such that 640 samples exactly match the 10-cycle
of the waveform. This changes the frequency resolution slightly, but the leakage
due to spreading of the fundamental frequency does not exist any more; however,
as mentioned in [23], it is important to note that the interpolation process may
distort the frequency content of the signal. But it is still much closer to the
exact solution than the standard FFT is, therefore it is used here for comparison
purposes. Interharmonic-1 is computed with group method of the IEC standard
[16]. The interharmonics computed with the proposed method, the method in
[23] and the standard FFT computation of 200ms windows are compared with
the reference interharmonics. The results are given in Table 3.3. It is observed
that the proposed method gives better results than the standard FFT algorithm,
computes interharmonics much closer to the reference values computed using the
algorithm given in [29]. Although for some frequencies, the proposed method
results are worse than the results of the method in [23], the proposed method has
some advantages over the method in [23]. In the proposed method, since FFT
algorithm is not applied for interharmonic analysis, there is no leakage eect
51
Table 3.3: Group computation of interharmonic-1 with the proposed
method, the method in [23], the standard FFT computation (FFT) and the
FFT+interpolation method on eld data. Error values are computed assuming
that the FFT+interpolation gives reference results.
Group Interharmonic-1
Fundamental FFT+Interpolation Proposed error Method in error Standard error
Freq (Hz) (Reference) Method (%) [23] (%) FFT (%)
49.90 0.163 0.162 0.85 0.162 0.61 0.178 8.94
49.93 0.160 0.160 0.25 0.149 6.83 0.236 47.50
50.02 0.131 0.135 3.21 0.129 1.51 0.148 12.98
50.12 0.162 0.167 2.79 0.165 2.96 0.188 12.58
50.15 0.165 0.168 2.06 0.166 0.60 0.175 6.06
problem. This also reduces the computational complexity. Furthermore, there
is no need for precise detection of the fundamental frequency, which means that
fundamental frequency deviation is not so critical in the proposed method. The
percentage dierence between the reference and the proposed method results are
given in boldface in the Table 3.3.
3.2.2 Discussion
In this section of the thesis, a Kalman ltering based method for interharmonic
computation is proposed for especially highly uctuating fundamental frequency
cases in the power system. Up to now, researches have shown that the system
frequency deviates signicantly due to the uctuating current demands of some
loads and disturbances or subsequent system transients in the power systems
especially when an interconnected system does not exist. If a standard window
length is used for the entire FFT analysis process, spectral leakage will occur
and even non-existing interharmonics in the signal content will appear at the
FFT output when the fundamental frequency deviates from the nominal value
[23]. The proposed method uses Kalman ltering to obtain the spectral de-
composition of the voltage waveform, which has some advantages over the FFT
based spectral decomposition methods. Since FFT algorithm is not applied,
the fundamental frequency deviation is not so critical, therefore no zero cross
detection algorithm is applied in this method. Only by determining the fun-
52
damental frequency as a state variable of the EKF, which is the rst Kalman
lter in the method, the fundamental frequency is computed. Also in [23], to
obtain the interharmonics, two FFT algorithms are applied which causes leak-
age eect problems in case of fundamental frequency deviation. However in the
proposed method, for a certain frequency resolution, suitable state variables are
determined for the Kalman lters and interharmonics of the waveform are ob-
tained. Higher frequency resolution can be preferred for more precise estimates,
although an increase in the frequency resolution has a cost of increasing the
matrix dimensions of the state equations.
The proposed method consists of two Kalman lters, which are an EKF and a
LKF. Initially an EKF is used which contains three state variables. These state
variables are in-phase component of waveform (x
1
), quadrature-phase compo-
nent of the waveform (x
2
) and nally fundamental frequency (x
3
). The output of
this lter gives the fundamental frequency and envelope of the voltage waveform
as shown in Figure 3.3. In the second part of the proposed method, a twenty-
state LKF is used. The DC component of the envelope of the waveform can
be considered as a sinusoidal at 0Hz. Therefore, for 5Hz frequency resolution
case, to obtain the subgroup and group computation of interharmonic-1, which
are the interharmonics between the fundamental and the second harmonic, the
voltage envelope is dened as the sum of ten sinusoidals at the corresponding
frequencies of 0, 5, 10, 15, 20, , 45Hz. Since for each of these sinusoidals, two
state variables of in-phase and quadrature-phase components are determined, a
twenty-state LKF model is obtained and the output of this LKF gives the spec-
tral amplitudes at the corresponding frequencies. In fact, to obtain the spectral
decomposition of the waveform, only a single Kalman lter can be used. In the
proposed method, the fundamental frequency is computed by the EKF. If it
is not necessary to obtain the fundamental frequency, only by using one LKF,
the spectral decomposition around the fundamental can be obtained. For this
case, a 38-state LKF model is used to obtain the interharmonics around the
fundamental. Equation (3.8) can be rewritten as given in
v(t) = V cos(2ft) +
N
i=1
V
i
4
{cos(2(f + f
fi
)t +
i
) + cos(2(f f
fi
)t +
i
)} .
(3.27)
53
As it is clearly observed from this equation that the eect of the modulation
appear as interharmonics at beat frequencies which are ff
fi
in the spectrum.
For 5Hz frequency resolution case, to obtain the interharmonics around the
fundamental, voltage waveform, v(t), can be considered as the sum of nineteen
sinusoidal waveforms at frequencies 5, 10, 15, , 90, 95Hz. For each of these
waveforms, two state variables of in-phase and quadrature-phase components
are determined and a 38-state LKF model is obtained. The output of the LKF
gives the spectral amplitudes at the interharmonics around the fundamental,
directly. Although the matrix dimensions of the state equations are increased,
the advantage of this approach is that in stead of using two types of Kalman
lters, only a single LKF is enough to obtain the same result.
This method provides good estimates of the interharmonics at frequencies around
the fundamental frequency. Since interharmonics are the main cause of icker,
the method can also be used for icker estimation algorithms, which are based
on spectral decomposition (such as [7] and [5]). In Section 3.3, such kind of a
icker estimation algorithm which is based on Kalman ltering is explained.
In this method, the window length is selected as 200ms as recommended in the
IEC Standard [16]. Therefore, for applications other than power quality, such
as real-time control and protection applications, which require response times
of the order of a couple of cycles, the method can be modied to use sliding
windows of 200ms overlapping one or two cycles, to be able to get updated
values every one or two cycles.
3.3 Flicker Evaluation Based on Kalman Filtering
Spectral decomposition obtained with Kalman lters is explained in detail in the
previous Section 3.2. In the proposed algorithm, V
i
/V values corresponding to
the icker frequencies are obtained by applying an EKF and a LKF consecutively
on voltage waveform using 10 cycle time windows which corresponds to window
length of 200ms for the 50Hz power system. Since Kalman lters are used to
estimate the signal spectrum, no leakage problem occurs, which is an important
54
advantage and novelty of this method.
Interharmonics are known to be the main cause of icker; therefore, the spectral
decomposition obtained by Kalman ltering, which is explained in Section 3.2,
is used here as a part of the proposed icker estimation algorithm. The icker
estimation approach in Chapter 2 is used; however, this time, the spectral de-
composition is obtained by Kalman ltering, not by applying FFT algorithm.
The block diagram of the proposed light icker estimation method, which is also
explained in [19], is given in Figure 3.6. In this diagram, the sampled input sig-
nal is denoted by v(t). Initially, in order to get rid of the harmonic eect, a low
pass lter (LPF) with a corner frequency at 75Hz is applied. Then for each time
window of ten cycles, the spectral decomposition is obtained by applying two
Kalman lters, EKF and LKF, consecutively. The EKF and LKF algorithms
are implemented based on [27].
LPF

v(t)
EKF
N rnek
Compute individual IFL
contributions according to IEC
61000-4-15
(from 0Hz to 25Hz)
sum(Si)
Statistical
Computation
S
P
st
A(t)cos(w
o
t+
)
EKF
N sample
A(t)

Spectral amplitudes
(from 0Hz to 25Hz)
S
i
LKF
N sample
sum(Si)
LPF
cutoff=75Hz
Figure 3.6: Block diagram of the proposed icker estimation method.
Since the voltage signal in EAF plants contains dierent icker frequencies, the
voltage uctuation is modeled as given in (2.2), which is again pointed out here
as,
v(t) = cos(2ft +
o
))
_
V +
N
i=1
V
i
2
cos(2f
fi
t +
i
)
_
. (3.28)
The signal model in (3.28) can also be modeled as a sinusoidal voltage waveform
with time varying amplitude A(t), where A(t) = V +
N
i=1
V
i
2
cos(2f
fi
t +
i
).
55
This signal model is shown as
v(t) = {cos(2ft +
o
)} A(t),
= cos(
o
)cos(2ft)A(t) sin(
o
)sin(2ft)A(t). (3.29)
In (3.29), A(t) represents the time varying amplitude of the input signal. In
order to obtain the spectral decomposition for the consecutive windows, EKF is
applied on the voltage waveform to obtain the envelope, A(t). The procedure of
obtaining the voltage envelope, A(t), by applying EKF is explained in Section 3.2
in detail. The state variables for this EKF algorithm are determined as, in-phase
component of v(t) (x
1
), quadrature-phase component of v(t) (x
2
), and nally
fundamental frequency of v(t) (x
3
).
Fundamental frequency deviation is frequently observed in EAF eld data [7, 19].
Since fundamental frequency is not a constant in the EKF algorithm, it is also
determined as the state variable, x
3
. Consequently, fundamental frequency of
the voltage waveform is also estimated successfully with the proposed spectral
decomposition method as explained in Section 3.2. Therefore, there is no extra
fundamental frequency computation algorithm in contrast to the FFT based
method given in Chapter 2, which is also proposed in [7]. The voltage signal is
a nonlinear function of the state variable, x
3
, therefore the state transition and
measurement equations are nonlinear functions of x
3
. In EKF algorithm state
transition and measurement equations are combined together in accordance with
the EKF model explained in [28] and the EKF model is realized. By applying
EKF on the voltage waveform, its envelope, A(t), is obtained. The linearized
state transition matrix, , and the linearized measurement vector, H, for the
EKF are given in Section 3.2 to explain the proposed spectral decomposition
method. The EKF algorithm is formed by using the linearized state transition
matrix, , and the linearized measurement vector, H. Then using the estimated
state variables of the state vector, which is the output of the EKF algorithm,
the envelope of the voltage waveform and the fundamental frequency variation
are obtained. The envelope of the voltage waveform is formed by using the state
variables x
1
and x
2
as
A(t) =
_
x
1
(t)
2
+ x
2
(t)
2
. (3.30)
56
The variation of the state variable x
3
gives the time-varying fundamental fre-
quency, which is represented as
f(t) = x
3
(t). (3.31)
Then, using a second Kalman lter model, which is LKF, spectral decomposition
of the voltage waveform is obtained. The input of the LKF is the envelope of
the voltage waveform, A(t), which is the output of EKF and contains dierent
icker frequency components. The mathematical representation of the envelope
of the voltage waveform, A(t), is
A(t) =
_
V +
N
i=1
V
i
2
cos(2f
fi
t +
i
)
_
. (3.32)
The envelope of the voltage waveform, A(t), is the measurement of the LKF
algorithm as given in the spectral decomposition method in Section 3.2. In the
FFT based method given in Chapter 2 [7], S is evaluated by using the frequency
spectrum of signal v(t) from f 25Hz to f +25Hz where f is the fundamental
frequency and 25Hz is the maximum icker frequency given in Table 2.1. How-
ever, in the Kalman ltering based method, the output of the EKF algorithm
is the envelope of the voltage uctuation, A(t), directly. Therefore, spectral
decomposition of A(t) is obtained by using a LKF for icker frequencies, f
fi
,
in the 0 25Hz frequency region. In equation (3.32), since V is constant, it
can be considered as a DC voltage signal. In the proposed method, A(t) is rep-
resented as the summation of six sinusoidal voltage uctuations at frequencies
0, 5, 10, 15, 20, 25Hz. 5Hz resolution is selected so that the results can be com-
pared with 5Hz resolution results of the proposed method in Chapter 2. Higher
frequency resolution can be preferred for more precise estimates, although fre-
quency resolution increase has a cost of increasing the matrix dimensions of the
state equations. For one of these sinusoidal voltage uctuations, the in-phase
and the quadrature-phase components, which are determined as the two state
variables for the corresponding voltage uctuation in the LKF model, are ob-
tained from sinusoidal expansion of the voltage uctuation as
V
i
2
cos(2f
fi
t +
i
) =
V
i
2
cos(
i
)cos(2f
fi
t)
V
i
2
sin(
i
)sin(2f
fi
t). (3.33)
57
In (3.33), states are dened as follows:
x
1
= in-phase component of the sinusoidal uctuation (
V
i
2
cos(
i
))
x
2
= quadrature-phase component of the sinusoidal uctuation (
V
i
2
sin(
i
))
i = 1, 2, , 6 represents the frequencies 0, 5, 10, 15, 20, 25Hz
The LKF model which is developed to obtain the spectral components of the
envelope of the signal, A(t), which contains six sinusoidal voltage uctuations
at dierent icker frequencies is given in equations (3.35) and (3.36).
Lets dene the state vector of the LKF, x
k
, as
x
k
=
_
x
1,1
x
2,1
x
1,2
x
2,2
x
1,6
x
2,6
_
T
k
(3.34)
where superscript T denes transpose of the vector given in brackets and
x
1,i
= in-phase component of the sinusoidal uctuation at icker frequency, f
fi
x
2,i
= quadrature-phase component of the sinusoidal uctuation at icker fre-
quency, f
fi
(f
f
= {0, 5, 10, 15, 20, 25})
Then the state transition equation for the twelve-state Kalman lter model is
given as
x
k+1
=
_
_
1 0 0
0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 1
_
_
x
k
+ w
k
. (3.35)
Since the state variables x
1,i
and x
2,i
represent the in-phase and quadrature-
phase components of the f
fi
voltage phasor, the measurement equation is formed
as
z
k
= x
k
+ v
k
, (3.36)
where the corresponding measurement vector, , is
=
cos(2f
f1
t) sin(2f
f1
t) cos(2f
f2
t) sin(2f
f2
t) cos(2f
f6
t) sin(2f
f6
t)
(3.37)
58
In this twelve-state LKF model, which is formed by equations (3.35) and (3.36),
w
k
represents the process noise vector and v
k
represents the measurement noise
vector. The spectral amplitude at icker frequency f
fi
is obtained by using the
corresponding state variables x
1,i
and x
2,i
, which are calculated for each time
window. The result
_
x
2
1,i
+ x
2
2,i
gives the spectral amplitude at icker frequency
f
fi
for the corresponding time window.
In the proposed method since the spectral amplitudes are obtained with 5Hz
resolution, icker frequencies in 0 25Hz frequency region, f
fi
, are determined
as 0, 5, 10, 15, 20 and 25Hz and twelve-state LKF model is obtained. However,
it is possible to obtain higher resolution for more accurate results, but this will
increase the number of states.
The spectral decomposition is obtained from 0Hz to 25Hz with 5Hz resolution
for each 10 cycle time window by applying EKF and LKF consecutively. The
V
i
V
value in equation (3.1) is obtained from this spectrum and S
i
is calculated
as given in (3.1). Finally, the instantaneous icker sensation, S, is obtained as
the summation of all S
i
.
3.4 Verication of the Proposed Method for Flicker Analysis
For verication of the proposed method, a digital realization of the IEC icker-
meter given in [9] is used to compare the response of the method to the actual
IEC ickermeter behavior with the proposed method. Both simulated and eld
data are used.
3.4.1 Verication with Simulated Data
Simulated data is produced by modulating the sine wave at 50Hz system fre-
quency as given in (2.2) with various icker frequencies and the corresponding
V/V values in Table 2.1 with a sampling frequency of 3200Hz. The data
length is 10 minutes. S is computed with the proposed method and the results
are compared with the results of the digital realization of the IEC ickermeter.
59
Table 3.4: Proposed light icker estimation algorithm response for sinusoidal
voltage uctuations
5.0 0.40 1.01 1.0 1.00 0.0
10.0 0.26 1.00 0.0 0.99 -1.0
12.0 0.31 0.99 -1.0 0.99 -1.0
15.0 0.43 0.98 -2.0 0.99 -1.0
20.0 0.70 0.99 -1.0 0.99 -1.0
25.0 1.04 1.02 2.00 1.02 2.0
3.4.1.1 Test 1: Single Flicker Frequency Case
The single sinusoidal icker frequency case is given in (2.1). It is expected that
S should be close to unity since Table 2.1 given in [1] is for unity S response to
sinusoidal voltage uctuations. The experiment is repeated with both 5Hz and
0.5Hz resolutions. The size of state vector is 6 in 5Hz resolution. However when
we increase the resolution rate to 0.5Hz, the size of state vector increases to 51
which means that resolution increase has a cost of increasing the number of state
equations. In case of 0.5Hz resolution, summation of all S
i
given in (3.1) from
0.5Hz to 25Hz with 0.5Hz increments are used to obtain S, whereas V/V
values corresponding to only 5, 10, 15, 20 and 25Hz are added up to obtain S
in case of 5Hz resolution. Comparison of the results with those of the IEC
ickermeter are given in Table 3.4 for both resolution values.
3.4.1.2 Test 2: Combinations of Dierent Flicker Frequencies Case
In this test, the signal model in (3.28) is used to produce a voltage waveform
at 3200Hz sampling frequency. The data length is again 10 minutes. This
simulated voltage waveform is shown as
v(t) =
_
1 +
1
2
10
100
cos
_
212t
3200
+

6
_
+
1
2
2
100
cos
_
25t
3200
+

3
__
cos
_
250t
3200
_
(3.38)
In Figure 3.7, simulated voltage waveform, v(t), for one window length (200ms)
60
0 50 100 150 200
1.5
1
0.5
0
0.5
1
1.5
time (msec)
M
a
g
n
i
t
u
d
e

(
p
.
u
)
Figure 3.7: Simulated voltage waveform, v(t), and the estimated envelope of
v(t) which is obtained by the EKF.
and the envelope of this voltage waveform, A(t), which is estimated by the rst
stage (EKF algorithm) is shown. As it is shown in Figure 3.7, the envelope of
the voltage waveform, A(t), is obtained successfully with the EKF. The short
term icker severity, P
st
, is equal to 22.53 with the proposed method, whereas
it is equal to 24.18 with the digital implementation of the IEC ickermeter.
3.4.1.3 Test 3: Homogeneity of the Flickermeter Response
The Standard IEC-61000-4-15 requires that the ickermeter response should
have the homogeneity property, i.e., whenever the V/V value is doubled or
halved, the resulting icker severity P
st
should also be doubled or halved re-
spectively [1]. Homogeneity of the proposed algorithm is checked by applying
twice and half of the V/V values with the corresponding icker frequencies
given in Table 3.4. It is observed that the proposed method estimates P
st
values
with negligible error, always less than 1.5% as given in Table 3.5 which means
that the homogeneity of the proposed algorithm holds. The error in this table is
dened as the percentage dierence between twice or half of the P
st
computed
61
Table 3.5: Homogeneity of the proposed light icker estimation algorithm
Flicker frequency (Hz) V/V % %
sinusoidal uctuation 230V/50Hz system P
st
Error
2.340 0.71
0.5 2 2.340 1.41 +0.70
0.5 2.340 0.36 -1.40
0.754 0.83
2.5 2 0.754 1.66 0.00
0.5 0.754 0.42 -1.20
0.398 0.69
5.0 2 0.398 1.38 0.00
0.5 0.398 0.35 -1.45
0.260 0.69
10.0 2 0.260 1.38 0.00
0.5 0.260 0.35 -1.45
0.312 0.70
12 2 0.312 1.39 +0.71
0.5 0.312 0.35 0.00
0.432 0.69
15 2 0.432 1.37 +0.72
0.5 0.432 0.35 -1.45
0.584 0.73
18 2 0.584 1.46 0.00
0.5 0.584 0.37 -1.37
0.700 0.69
20 2 0.700 1.37 +0.72
0.5 0.700 0.35 -1.45
1.042 0.70
25 2 1.042 1.40 0.00
0.5 1.042 0.35 0.00
and the P
st
obtained by the proposed method when V/V is doubled or halved.
3.4.2 Verication on AC Arc Furnace Data
Experiments on eld data obtained at three dierent EAF plants are used to
verify the eectiveness of the proposed algorithm in estimating the light icker.
The proposed light icker estimation algorithm is used with 5Hz frequency
resolution (200ms windows). The data is obtained during the eld measurements
carried out through the National Power Quality Project of Turkey [22, 26]. EAF
voltages are collected at a sampling rate of 3200Hz at the MV side of the EAF
62
transformer [25]. Rms voltages collected at the MV side of the EAF transform-
ers at those three plants are given with the instantaneous icker sensation, S,
values which is obtained from both the digital implementation of the IEC ick-
ermeter and the proposed light icker estimation algorithm in Figures 3.8, 3.9,
and 3.10. In these gures, (a) shows the rms voltages with the P
st
computed
intervals, (b) shows the frequency variation, and (c) shows S computed by both
the proposed method and the IEC ickermeter, for comparison purposes. The
frequency variation is given to show that the proposed method gives good esti-
mates irrespective of the frequency variation. It is observed that the proposed
method gives good estimates of the instantaneous icker sensation, S. More
detailed views of S for the rst two intervals for Figures 3.8, 3.9, and 3.10 are
given in Figures 3.11, 3.12, and 3.13, respectively, for better comparison. To
give a quantitative comparison of the algorithm, short term icker severity, P
st
is computed for the intervals given in the gures and the results are compared
with both the results of the digital realization of the IEC ickermeter and the
rst proposed method which is also explained in [7] as given in Tables 3.6, 3.7
and 3.8. The results indicate good agreement between the IEC ickermeter
and the proposed method which means that proposed method is successful in
estimating the light icker.
63
P
st
icker sensation, S.
st
in Fig 3.8 with the proposed method, method in [7] and the digital realization
of the IEC ickermeter
INTERVALS IEC Flickermeter Method in [7] Err % Proposed Mtd. Err %
INTERVAL 1 15.4958 15.4131 +0.53 15.5314 -0.23
INTERVAL 2 9.5497 9.6166 -0.70 9.5722 -0.24
INTERVAL 3 13.2904 13.5436 -1.91 13.2257 +0.49
INTERVAL 4 2.6938 2.6922 +0.06 2.7760 -3.05
INTERVAL 5 1.8163 1.9905 -9.59 1.9377 -6.68
64
P
st
icker sensation, S.
st
INTERVAL 1 22.3209 23.3268 -4.51 22.0043 +1.42
INTERVAL 2 21.4275 22.5856 -5.40 22.1809 -3.52
INTERVAL 3 11.5852 11.9070 -2.78 11.6648 -0.69
INTERVAL 4 9.2761 9.4001 -1.34 9.0635 +2.29
INTERVAL 5 10.6356 10.9129 -2.61 10.8473 -1.99
65
P
st
icker sensation, S.
st
INTERVAL 1 11.6423 11.7285 -0.74 11.9820 -2.92
INTERVAL 2 16.0384 15.6946 +2.14 15.9033 +0.84
INTERVAL 3 9.0006 9.2464 -2.73 9.1137 -1.26
INTERVAL 4 8.2767 8.8416 -6.83 8.7478 -5.69
INTERVAL 5 6.4816 7.4208 -14.49 5.9577 +8.08
66
67
The calculation procedures are analyzed on a personal computer (PentiumIV,
2.20GHz). All algorithms are implemented using MATLAB running on Win-
dows XP. The total computation time for a 1min input data is 75s. However,
as mentioned in the spectral decomposition method part in Section 3.2, this
computation time can be reduced much more by using a faster compiling envi-
ronment such as C and a faster processor. The code can also be optimized for
ecient computation.
The accuracy obtained on eld data are comparable with the results given in
[5] and in [2]. The mean error is 2.14%, 1.98%, and 3.76% for Plant-1, Plant-2
and Plant-3, respectively. In [2] mean error is around 6%. It varies between
0.44% and 5.94% in [5]. In [7] the mean error varies between 2.56% and 5.39%
for dierent EAF plants.
Fundamental frequency for the three EAF plants are calculated by using EKF.
Data length of each arc furnace data is chosen as 10 minutes and for each of them,
histogram of the computed fundamental frequencies are obtained. The his-
togram of fundamental frequencies for Plant-1, Plant-2, and Plant-3 are shown
in Figures 3.14, 3.15, and 3.16, respectively.
68
49.8 49.9 50 50.1 50.2 50.3
0
1
2
3
4
5
6
7
x 10
4
Histogram of Fundamental Frequency of Plant1
Freq (Hz)
Figure 3.14: Histogram of fundamental frequency of Plant-1.
49.8 49.9 50 50.1 50.2 50.3
0
1
2
3
4
5
6
7
x 10
4
Freq (Hz)
69
49.8 49.9 50 50.1 50.2 50.3
0
1
2
3
4
5
6
7
x 10
4
Freq (Hz)
Since the histogram of fundamental frequency of each eld data is similar to a
Gaussian distribution form, we can again say that Kalman application provides
quite satisfactory results.
70
CHAPTER 4
CONCLUSIONS
In this thesis, new ickermeter applications are proposed as alternatives to the
IEC-ickermeter. Two dierent novel spectral-decomposition-based methods are
proposed to estimate the light icker of EAF plants, where the system frequency
deviates signicantly. The frequency deviation of the EAF plants is inside the
range from 49.7Hz to 50.3Hz, where the nominal system frequency is 50Hz. In
the rst proposed method, spectral decomposition is obtained by utilizing the
DFT algorithm on the voltage waveform directly using 10-cycle time windows.
In the second method, spectral decomposition is obtained by utilizing two types
of Kalman lters (EKF and LKF) consecutively. Again 10-cycle time windows
are used, which coincides with the IEC harmonic and interharmonic computation
standard [16].
For both of the proposed methods, instantaneous icker sensitivity is obtained
from the addition of contribution of each icker-creating spectral component in
the 25Hz neighborhood of the fundamental frequency. In the rst method, leak-
age eect of the DFT algorithm in the case of deviated fundamental frequency is
corrected by vector subtraction of the leakage DFT components from the DFT
components of the original voltage waveform and the corrected spectral compo-
nents are obtained. The leakage components are obtained from a pure sinusoidal
waveform synthesized at the exact fundamental frequency. However, in the sec-
ond method, the EKF algorithm computes the envelope of the voltage waveform
and the fundamental frequency is eliminated. Since no DFT is involved, there
is no risk of spectral leakage due to the fundamental frequency deviation. The
second approach does not require a spectral correction procedure; which is an
71
advantage of this method.
The algorithms are veried both on simulated and eld data obtained from three
dierent EAF plants. The comparison with the digital realization of the IEC
ickermeter shows that both of the methods give satisfactory estimations of the
light icker with low computational complexity.
In particular, for the rst approach, which is the spectral decomposition based
one for light icker evaluation of electric arc furnaces, the following should be
noted:
The proposed approach gives good estimates of the instantaneous icker
sensation of the EAFs when compared with the digital realization of the
IEC ickermeter. Hence the short term icker severity, P
st
, which is com-
puted from the instantaneous icker sensation is also estimated success-
fully. The proposed method is especially suitable to estimate the light
icker at the MV busbar of EAFs, if the short circuit MVA of the busbar
is not high enough to keep the system frequency deviation inside a tolerable
range. When a xed FFT window (typically 200ms long) is used without
any spectral correction, even a deviation of 0.01Hz causes signicant error
in interharmonic estimation and hence the icker severity.
The proposed method has low computational complexity. All computa-
tions are simple vector operations achieved on spectral samples of the FFT
of the voltage waveform in the neighborhood of the fundamental frequency.
In most of the power quality (PQ) analyzers, FFT of the voltage waveform
is already obtained for harmonic analysis. Moreover, the same frequency
resolution and the FFT size are used for the harmonic computation are
as recommended in the IEC standard [16]. Therefore, computation of the
icker will not load an extra burden if FFT is already being used in a PQ
analyzer.
The use of a time window of 10 cycles as recommended in the IEC standard
does not give good results when the fundamental frequency deviates, espe-
cially during the boring and melting phases of the EAF operation, where
72
high currents are drawn from the busbar. This is because of the leakage
eect of the FFT process when the window size is not an integer multi-
ple of the fundamental period. Therefore a leakage correction process is
employed in the proposed approach. This correction leads to very success-
ful estimates of both the instantaneous icker sensation, S, and the short
term icker severity, P
st
. However, for this correction to be successful, an
accurate estimation of the fundamental frequency is essential.
Related to the previous item: Once the fundamental frequency is esti-
mated, another method to correct the FFT leakage would be to adjust the
window length and recompute the FFT; however, this will result in window
sizes which are not integer multiples of 2
n
, where n is a positive integer,
and the eciency of the FFT algorithm will be reduced. This problem
can be solved by changing the sampling frequency of the device; however,
this would require re-sampling the signal, which is not preferred for on-
line computations. Sampling frequency can also be changed by applying
interpolation on the already-sampled signal, but this will always contain
some error of the original signal due to the interpolation procedure. The
proposed method helps to avoid these problems.
The proposed method provides a good understanding of the eect of fre-
quency spectrum of the voltage on the icker sensation of the human eye
and the brain.
In the proposed method, the existence of the fundamental frequency is
considered without any harmonics. In case of harmonics existing in the
spectrum, interharmonic frequencies around the existing harmonics should
also be considered. For example, in the existence of the second harmonic,
the voltage waveform will be given as
v(t) = [V sin(2ft) + V
2
sin(22ft)]
_
1 +
V
2V
sin(2f
f
t)
_
(4.1)
where V
2
is the second harmonic amplitude. In this case, spectral compo-
nents will appear not only at ff
f
, but also at 2ff
f
, with amplitudes
(V )/4 and (V )V
2
/4V respectively. The components at 2ff
f
will
be negligible when (V )V
2
/4V ratio is negligible compared with V/4,
73
which is usually the case. However, if it is known that the foregoing EAF
voltage contains signicant amount of the harmonics compared to the fun-
damental, then the eects of the harmonics should also be considered in
the instantaneous icker sensation computation in the summation given
in (2.10) with appropriate coecients.
In addition, the following should be noted about the second proposed approach,
Kalman-ltering based approach for light icker evaluation of electric arc fur-
naces :
This proposed approach gives good estimates of the instantaneous icker
sensation when compared with the digital realization of the IEC icker-
meter. Hence the short term icker severity, P
st
, which is computed from
the instantaneous icker sensation is also estimated successfully. The pro-
posed method is especially suitable to estimate the light icker at the MV
busbar of EAFs, if the short circuit MVA of the busbar is not high enough
to keep the system frequency deviation inside a tolerable range.
The proposed method has low computational complexity. Only two types
of Kalman lters with easy implementations are used and the icker caus-
ing interharmonic frequencies are obtained. Moreover, the frequency res-
olution and the window size in the harmonic computation are used as
recommended in the IEC standard [16].
Arc furnace data is not stationary; therefore, time window of longer than
10 cycles would not be a good choice. All the results in the proposed
method are obtained with 5Hz resolution, but it is possible to obtain a
higher resolution by increasing the number of state variables.
The proposed Kalman-ltering based method consists of two Kalman l-
ters, which are an EKF and a LKF. Initially an EKF is used. The output
of this lter gives the fundamental frequency and envelope of the voltage
waveform. In second part of the proposed method, a twelve-state LKF is
used to obtain the spectral amplitudes at the corresponding icker frequen-
cies of 0, 5, 10, , 25Hz using the voltage envelope. In fact, to obtain the
74
spectral decomposition of the waveform, all these can also be done by us-
ing only a single Kalman lter. In the proposed icker estimation method,
the fundamental frequency is computed by the EKF. If it is not necessary
to obtain the fundamental frequency, only by using one LKF, the spectral
decomposition around the fundamental can be obtained. For this case,
a 22-state LKF model is used to obtain the interharmonics around the
fundamental. Equation (3.8) can be rewritten as given in
v(t) = V cos(2ft)+
N
i=1
V
i
4
{cos(2(f + f
fi
)t +
i
) + cos(2(f f
fi
)t +
i
)} .
(4.2)
As it is clearly observed from this equation that the eect of the modula-
tion appear as interharmonics at beat frequencies which are ff
fi
in the
spectrum. For 5Hz frequency resolution case, to obtain the interharmon-
ics around the fundamental, voltage waveform, v(t), can be considered as
the sum of eleven sinusoidal waveforms at frequencies 25, 30, , 70, 75Hz.
For each of these waveforms, two state variables of in-phase and quadrature-
phase components are determined and a 22-state LKF model is obtained.
The output of the LKF gives the spectral amplitudes at the interharmon-
ics around the fundamental, directly. Although the matrix dimensions of
the state equations are increased, the advantage of this approach is that
in stead of using two types of Kalman lters, only a single LKF is enough
to obtain the same result.
Another advantage of the proposed algorithm is that the selection of exact
window size is not so critical. When the fundamental frequency deviates
from its nominal value, the constant window size causes a considerably
signicant error in DFT-based algorithms. The proposed algorithm re-
sults, on the other hand, are not aected that much by the window size
change as it can be observed from Table 4.1. The results in Table 4.1 are
obtained by simulating a voltage waveform with single sinusoidal icker
frequency as given in (2.1). In this voltage waveform, V/V is 0.260 and
f
f
is 10Hz, which gives unity S response with the proposed algorithm as
it is given in Table 3.4. Sampling rate is chosen as 3200Hz and resolution
rate is 5Hz, which means that there should be 640 samples per window
75
Table 4.1: The Eect of Changing the Number of Samples per Window to
Instantaneous Flicker Sensation, S
Number of Samples S Number of Samples S
640 1
635 0.99 645 0.99
630 0.99 650 0.98
625 0.98 655 0.98
620 0.98 660 0.97
when the fundamental frequency is exactly 50Hz. Table 4.1 shows the S
responses for dierent number of samples per window. The results show
that the fundamental frequency deviation is not so critical for the proposed
method. S is estimated as close as 97% to the actual S value, even if the
window size is changed from 640 samples to 660 samples.
When the Kalman ltering based approach of this research work is compared
with the rst spectral decomposition based approach on icker computation,
several points should be emphasized:
The Kalman ltering based approach does not suer from fundamental fre-
quency deviation and hence does not require the spectral correction proce-
dure applied in the rst spectral decomposition based approach. Moreover,
the fundamental frequency is computed inherently in the rst step of the
algorithm. The method does not include any fundamental frequency detec-
tion algorithm such as zero cross detection explained in the rst proposed
method [7] for precise detection of the fundamental frequency. However, if
it is necessary to compute the fundamental frequency exactly, only by de-
termining fundamental frequency as a state variable in the EKF algorithm,
it is possible to obtain it at an intermediate step of the algorithm.
The Kalman ltering based approach uses 10 cycle windows for analysis.
It is possible to reduce the window size for less stationary voltage signals,
such as the case in EAF plants, without losing the frequency analysis
resolution, at the expense of less accurate frequency estimates. However,
in the DFT-based approach, reducing the window size means reducing the
76
frequency resolution and hence the algorithm accuracy.
The rst method is faster than the second one. Therefore, a compromise
should be made between the computation time and the accuracy of com-
putation.
77
REFERENCES
[1] IEC 61000-4-15: Testing and Measurement Techniques - Flickermeter -
Functional and Design Specications, 2002.
[2] C.J. Wu, Y.J. Chen, A Novel Algorithm for Precise Voltage Flicker Cal-
culation by Using Instantaneous Voltage Vector, IEEE Transactions on
Power Delivery, vol. 21, no. 3, pp. 15411548, 2006.
[3] S.J. Huang, C.W. Lu, Enhancement of Digital Equivalent Voltage Flicker
Measurement via Continuous Wavelet Transform, IEEE Transactions on
[4] M.T. Chen, Digital Algorithms for Measurement of Voltage Flicker, IEE
Proceedings- Generation, Transmission and Distribution, vol. 144, no. 2,
pp. 175180, 1997.
[5] A. Hernandez, J.G. Mayordomo, R. Asensi, et.al, A New Frequency Ap-
proach for Flicker Evaluation of Arc Furnaces, IEEE Transactions on
[6] C.J. Wu, T.H. Fu, Eective Voltage Flicker Calculation Algorithm Using
Indirect Demodulation Method, IEE Proceedings- Generation, Transmis-
sion and Distribution, vol. 150, issue 4, pp. 493500, 2003.
[7] N. Kose, O. Salor, New Spectral Decomposition Based Approach for
Flicker Evaluation of Electric Arc Furnaces, IET Generation, Transmis-
sion, Distribution Journal, vol. 3, issue 4, pp. 393411, 2009.
[8] C.W. Lu, S.J. Haung, C.L. Huang, Flicker Characteristic Estimation of an
AC Arc Furnace, Electric Power Systems Research, Elsevier, vol. 54, no.
2, pp. 121130, 2000.
[9] T. Keppler, N.R. Watson, S. Chen, J. Arrilaga, Digital ickermeter real-
isations in the time and frequency domains, Proceedings of Australasian
Power Engineering Conference, Perth Australia, Sept. 2001.
[10] M. Gol, O. Salor, B. Alboyaci, B. Mutluer, I. Cadirci, M. Ermis, A new
eld-data based EAF model for power quality studies, Proc. of the IEEE
Industry Applications Society Annual Meeting, Canada, 2008.
[11] R. Thottungal, P. Anbalagan, T. Mohanaprakash, et.al., Power System
Stabilization in Multi Area System by Using HVDC Link, Academic Open
Internet Journal (www.acadjournal.com), vol. 19, 2006.
[12] X.J. Jiang, Z.Y. Tang, Z.Y. Dong, T.K. Saha, An Advanced Method for
Eliminating Impacts from Frequency Deviations in Power System Signal
78
Processing, Electric Power Systems Research, Elsevier, vol. 67, no. 3, pp.
177184, 2003.
[13] C.S. Chen, H.J. Chuang, C.T. Hu, S.M. Tseng, Mitigation of Voltage
Fluctuation for an Industrial Customer with Arc Furnace, Proceedings of
IEEE Power Engineering Society Summer Meeting, vol. 3, pp. 16101615,
2001.
[14] V. Terzija, V. Stanojevic, Robust Estimation Algorithm for Power Compo-
nents Measurement, Proceedings of IEEE PES Power Systems Conference
and Exposition, pp. 22292235, 2006.
[15] Y. Liu, A US-Wide Power Systems Frequency Monitoring Network, Pro-
ceedings of IEEE PES Power Systems Conference and Exposition, pp. 159
166, 2006.
[16] IEC 61000-4-7: Testing and Measurement Techniques - General Guide
on Harmonics and Interharmonics Measurements and Instrumentation for
Power Supply Systems and Equipment Connected Thereto, 2003.
[17] N. Kose, O. Salor, A New Frequency Domain Approach for Light Flicker
Evaluation of Power Systems, I2MTC 2009, International Instrumentation
and Measurement Conference, Singapore, May 5-7 2009.
[18] N. Kose, O. Salor, K. Leblebicioglu, Interharmonics Analysis of Power
Signals with Fundamental Frequency Deviation Using Kalman Filtering ,
submitted to Electric Power Systems Research, Elsevier, July 24 2009.
[19] N. Kose, O. Salor, K. Leblebicioglu, A Kalman Filtering Based Approach
for Light Flicker Evaluation of Power Systems, submitted to IET Gener-
ation, Distrubution and Transmission Journal, March 05 2009.
[20] J.A. Pomilio, S.M. Deckmann, Flicker Produced by Harmonics Modula-
tion, IEEE Transactions on Power Delivery, vol. 18, no. 2, pp. 387392,
2003.
[21] T. Keppler, N.R. Watson, J. Arrilaga, S. Chen, Theoretical Assessment
of Light Flicker Caused by Sub- and Interharmonic Frequencies, IEEE
Transactions on Power Delivery, vol. 18, no. 1, pp. 329333, 2003.
[22] T. Demirci, A. Kalaycioglu, O. Salor, et. al., National PQ Monitoing Net-
work for Turkish Electricity Transmission System, Proceedings of Instru-
mentation and Measurement Technology Conference, Poland, May 1-3 2007.
[23] O. Salor, Spectral Correction Based Method for Interharmonics Analysis
of Power Signals with Fundamental Frequency Deviation, Electric Power
Systems Research, Elsevier, vol. 79, pp. 10251031, 2009.
[24] M.B. Duric, Z.R. Durisic, Frequency measurement in power networks in
the presence of harmonics using Fourier and zero crossing technique, IEEE
Proc. Power Tech., pp. 16, 2005.
79
[25] E. Ozdemirci, Y. Akkaya, B. Boyrazoglu, et. al., Mobile Monitoring Sys-
tem to Take PQ Snapshots of Turkish Electricity Transmission System,
Proceedings of Instrumentation and Measurement Technology Conference,
Poland, May 1-3 2007.
[26] G uc Kalitesi Milli Projesi, http://www.guckalitesi.gen.tr, last visited on
August 2009.
[27] G. Welch, G. Bishop, An Introduction to the Kalman Filter Technical
Report, University of North Carolina at Chapel Hill, 2006
[28] A. Girgis, J. Stephens, E. Makram, Measurement and Prediction of Volt-
age Flicker Magnitude and Frequency, IEEE Transactions on Power De-
livery, vol. 10, no.3, 2005.
[29] G.W. Chang, C.I. Chen, Y.J. Liu, M.C. Wu, Measuring power system
harmonics and interharmonics by an improved fast Fourier transform-based
algorithm, IET Generation, Transmission and Distribution Journal, vol.
2, no. 2, pp. 192201, 2008.
80
APPENDIX A
IEC FLICKERMETER RESPONSE FOR
RECTANGULAR FLUCTUATIONS
The IEC ickermeter response for rectangular uctuations for 230V lamp and
50Hz system is given in Table A.1.
Table A.1: Normalized Flickermeter Response for Rectangular Voltage Fluctu-
ations, 230V/50Hz system - IEC Standard 61000-4-15
Hz Voltage Fluctuation % Hz Voltage Fluctuation %
0.5 0.514 10.0 0.205
1.0 0.471 10.5 0.213
1.5 0.432 11.0 0.223
2.0 0.401 11.5 0.234
2.5 0.374 12.0 0.246
3.0 0.355 13.0 0.275
3.5 0.345 14.0 0.308
4.0 0.333 15.0 0.344
4.5 0.316 16.0 0.376
5.0 0.293 17.0 0.413
5.5 0.269 18.0 0.452
6.0 0.249 19.0 0.498
6.5 0.231 20.0 0.546
7.0 0.217 21.0 0.586
7.5 0.207 22.0 0.604
8.0 0.201 23.0 0.680
8.8 0.199 24.0 0.743
9.5 0.200 33.3 1.670
81

Index

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Index

Încărcat de

Drepturi de autor:

Formate disponibile

LIGHT FLICKER EVALUATION OF ELECTRIC ARC FURNACES BASED

ON NOVEL SIGNAL PROCESSING ALGORITHMS

OSE in partial fulllment of the requirements for

ITAK-UZAY; Electrical and Elec-

ITMALARI KULLANARAK ELEKTR

ITAK-UZAY Power Electronics Group. I would like to thank

ITAK for full nancial support of the project.

S-ar putea să vă placă și